P
Platformization
Guest
One week after launch, GPT-6 Luna and Sol are already running in clouds, code editors, invoice pipelines, trading agents, whisky classifiers and CI bots. Here are the real ones, sorted by what you are trying to build.
OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026. Sol costs $2 per million input tokens and $10 per million output tokens, and it is aimed at coding and agent work. Luna costs $0.10 in and $0.50 out, and it is aimed at high-volume extraction, summarizing, classification and routing. Both are about half the price of the GPT-5.6 models they replace, and cached input is 90% off (OpenAI, VentureBeat).
Every example below links to something you can check: a merged pull request, a release, a platform's docs, or a write-up with numbers. Entries marked (PR) are pull requests that were not yet merged when this was written. The last section collects what went wrong, because a list of only wins is not much use when you're deciding what to build.
The pattern across these projects: Luna goes on work that is high-volume and has a clear right answer, and Sol goes on coding, review and anything that needs judgment.
You don't need an OpenAI account to use GPT-6. Within a week it was available through every major cloud except Google, and through most model gateways.
1. OpenAI API —
2. ChatGPT Work and desktop app — paid plans get both models in Work and Codex; free users get Luna, but only in the desktop app. Notebookcheck
3. Microsoft Foundry (Azure) — Astra, Sol and Luna in 28 regions, with Luna pitched for extraction, routing and classification. Azure blog
4. Amazon Bedrock — Sol for coding and data analysis, Luna for extraction and Q&A at scale, with prompt caching and optional zero data retention. AWS blog
5. Snowflake Cortex AI — call Sol or Luna straight from SQL through Cortex AI Functions (public preview). Snowflake
6. Databricks Foundation Model APIs — both models listed as supported. Databricks docs
7. Microsoft 365 Copilot and Copilot Studio — Sol (not Luna) in Word, Excel, PowerPoint and Copilot Studio, next to Claude Opus 5.5. Windows Report
8. OpenRouter — both models, routed across OpenAI, Azure and Bedrock. In week one Codex ran 166B Sol tokens through it, Hermes Agent 82.5B and pi 53.5B. Sol · Luna
9. Vercel AI Gateway — same-day support; Sol for complex agent work, Luna as the cheaper option. Changelog
10. Cloudflare AI — both models; Sol goes to $4/$15 for long-context requests. Sol · Luna
11. LiteLLM — works from launch day on v1.101.0+, with pricing in its cost map. LiteLLM blog
12. Requesty — Standard and Flex tiers, with Luna pitched for structured extraction. Requesty
13. TrueFoundry AI Gateway — suggests routing each request to Sol or Luna by complexity and claims 50–70% savings. TrueFoundry
14. Portkey — pricing and capability entries in its open model catalog. PR #947
15. CometAPI — resells Sol input at $1.60 per million, 20% under OpenAI's price. CometAPI
16. EvoLink — Luna through its Chat Completions and Responses endpoints. EvoLink
17. NEAR AI Cloud — fixed tool calling so both models work with reasoning turned off. PR #1121
18. Pollinations — open generative-AI API added both as paid-only models. PR #15312
19. OpenGradient TEE Gateway — both models served from a hardware-secured enclave. PR #171
20. ClawRouter — multi-provider router added endpoints, limits and pricing for both. PR #247
21. Coder's xum — routing layer adds both, keeping old model IDs working. (PR) PR #4348
22. Windmill Luna Gateway — an OpenAI-compatible gateway for Windmill moved its default from GPT-5.6 Luna to GPT-6 Luna. PR #13
23. Felo AI — Sol in Felo Search and the Felo API, for batch summarizing and extraction. Felo
24. Emergent — Sol available through its single "Universal Key" for app building. Emergent
25. Puter.js — free access to both models from the browser with no API key; Puter's cost model puts an 80-turn coding session at $2.32 on Sol and $0.116 on Luna, with caching saving more than the rate cut. Puter
26. Make — the no-code automation platform added both to its AI Toolkit, OpenAI app and AI Agents, suggesting Luna for categorizing and routing support requests and Sol for turning bug reports into fix drafts. Make community
The fastest adoption was in coding tools. Most made Sol the default for writing code and kept Luna for small, fast tasks.
27. OpenAI Codex — Sol and Luna added to the model catalog, with migration paths from GPT-5.x. Simon Willison made Sol his Codex default. PR #47332
28. GitHub Copilot — Sol and Luna in VS Code, Visual Studio, JetBrains, Xcode, Eclipse, the CLI and github.com. Changelog
29. Zed — both models for users who bring their own OpenAI key. PR #64628
30. Warp — each model at five reasoning levels, low to max. Warp docs
31. Factory (Droid CLI) — Luna is billed at 0.04× a standard request and Sol at 0.8×, so Luna is the cheap default for routine edits. Factory docs
32. Kilo Code — both models; Kilo estimates about $0.18 to have Sol analyze a 10,000-line codebase. Sol · Luna
33. OpenCode — both models through ChatGPT/Codex sign-in. PR #50704
34. Pi (pi.dev) — terminal coding agent offers Sol through Copilot, Azure and OpenAI. Pi
35. Codex plugin for Claude Code — OpenAI's plugin made Sol the default for new tasks. PR #784
36. gptme — terminal agent added both to its model registry. PR #3914
37. maki — Lua-scriptable coding agent added both for ChatGPT Coding Plan users (a fork, makima, did the same). PR #1088 · makima #145
38. Overdeck — multi-agent orchestrator for coding assistants registered both in every picker. PR #4234
39. Phoenix IDE — both models on direct OpenAI and Codex routes. PR #801
40. microcodex — ultra-light terminal agent switched its default from GPT-5.6 Sol to GPT-6 Sol. PR #23
41. ChatGPT for VS Code (community fork) — added both, with Sol as the default. Release
42. Background Agents — open-source background coding agents added both to their Slack, Linear and GitHub bots. PR #2028
43. opencodex — ranks GPT-6 ahead of GPT-5.6 Sol across providers. (PR) PR #5584
44. makeitworkcloud Helm charts — moved 9 production OpenCode agents from GPT-5.6 Terra to GPT-6 Sol. PR #114
45. A developer's dotfiles — Pi coding-agent routes moved to GPT-6, with GPT-5.6 Terra kept as a fallback. PR #31
46. Paperclip — open-source agent manager adds Sol to its Codex adapter, with reasoning from low to ultra. (PR) PR #13830
47. GItS — bumped its Codex CLI so Sol is the default. PR #655
48. OpenClaw Code Agent — allows Luna and prices both in its cost estimator. (PR) PR #479
This is where Luna has the clearest numbers so far: cheaper than the model it replaced, and in several cases more accurate too.
49. Facturear (invoice extraction) — on 22 real invoices, Luna got 93.2% of fields right vs 82.5% for gpt-4o. Tax IDs were right on 21 of 22 vs 10 of 22, median time fell from 36–44s to 15.5s, and cost fell from $0.021–0.034 to about $0.0044 per invoice. PR #37
50. EICR (live extraction backend) — moving from GPT-5.6 Luna cut median latency from 4,494 to 3,354 ms and cost by 55%, with all 13 test fixtures matching. PR #221
51. reflection (claim extraction) — 56 of 56 outputs valid and 58% cheaper than GPT-5.6 Luna. Precision rose from 57% to 74% while recall fell from 85% to 73%. PR #28
52. which-glp (post extraction) — Luna was at least as accurate as gpt-5-nano, which misread "8 lbs" once, at about 1.4× nano's cost. PR #214
53. Catholic theology search (RAG) — swapping gpt-5.4-mini for Luna raised answer quality from 4.19 to 4.67 out of 5 and cut unsupported claims from 16.7% to 1.7%, at about 12% lower cost. PR #114
54. Voicebook — a local-first document reader with text-to-speech made Luna its default, replacing the GPT-5.6 models. PR #159
55. The Plow Times — a research-paper generator that delivers sourced PDFs moved from Claude Opus 5 to GPT-6 Sol via OpenRouter, using the 1M-token context. PR #132
56. Sure (personal finance) — adds Sol for chat, vision and PDF processing of statements. (PR) PR #3771
57. reviewer2 (academic paper reviewer) — fixed failing GPT-6 requests and added GPT-6 pricing, moving its docs off gpt-5-mini. PR #3
58. PaperPilot (Zotero research tool) — Sol and Luna added to its Codex model picker for working through papers. PR #76
Put something in a bucket, pick a label, decide where a request goes: this is the work Luna was priced for.
59. Microsoft openaivec — Microsoft's SDK for running OpenAI models over Spark and DuckDB made Luna its new default, replacing gpt-4.1-mini for classification, extraction and translation. Release v2.7.0
60. itemize (Monarch Money transactions) — Luna categorizes retail transactions, replacing GPT-5.6 Luna at $0.10/$0.50 vs $0.20/$1.20 per million tokens. PR #97
61. Peated (whisky-bottle photos) — David Cramer's classifier matched GPT-5.6 Luna at 79 of 105 correct, and cost fell 31% from $0.32 to $0.22, though runs got slower, from 22.5s to 31.4s. PR #1311
62. phototag (Lightroom photo tagging) — default vision model moves from gpt-4o-mini to Luna. (PR) PR #1
63. Telegram bot — Luna decides whether the bot should reply and handles web search, replacing GPT-5.6 Luna. PR #75
64. jev-issue-router — recommends model and reasoning settings for each GitHub issue; added Sol and Luna in place of GPT-5.6 Terra. PR #5
65. Bühler Group internal chat — the industrial company's chat app makes Sol the default and drops to Luna when spending caps are hit. (PR) PR #182
66. kipp workflows — replaced GPT-5.6 Luna with the
The most common move in week one: make Luna the default model in a product, and keep Sol for the users or tasks that need more.
67. Dust — the enterprise agent platform made Luna the default for its built-in agents, citing half the input price and 2.4× cheaper output than GPT-5.6 Luna. PR #33031 · docs
68. Perplexity — Sol is the default for the "Light" effort setting, including its Computer agent. Crypto Briefing
69. Datasette Agent — Simon Willison moved the agent.datasette.io demo to Luna and calls it "fast and competent" at SQL queries and HTML/JS apps. simonwillison.net
70. Hermes Agent (Nous Research) — swaps its GPT-5.6 tiers for GPT-6 and is the #2 app by Sol tokens on OpenRouter. (PR) PR #119410
71. chatniuexa (chatbot builder) — Luna replaces gpt-5.4-mini as the default, and bots that were on flagship models move to Sol; the team says Luna is "~7-9x cheaper". PR #2
72. VYNN AI stock analyst — Luna replaces gpt-5.4-mini for chat and gpt-4o-mini for analysis; a full analysis ran with no API errors. PR #22
73. TradingAgents — the multi-agent trading framework uses Sol for analysis, a scheduled writer/judge, YouTube synthesis and portfolio decisions. PR #288
74. TollChat (Northern Virginia commuter finance agent) — the app, its simulated users and its judges all moved to Luna; a 600-trial baseline passed 68.97% of cases for $1.72 total. PR #590
75. In-game character agents (CS3216 project) — p90 latency fell from 3.91s to 2.43s and cost per call from $0.00094 to $0.00041, with 30 of 30 valid outputs and 5 of 5 safe replies to hostile prompts. PR #138
76. Cadencia-AI (goal planner) — moved from DeepSeek to Luna in production; a goal run went from 65,472 to 23,584 micro-dollars. PR #30
77. Open Science Assistant — default set to Luna on Bedrock; three community assistants answered live with citations. PR #529
78. Stately Agent — David Khourshid's state-machine agent library made Luna its demo default, with Sol for one script; 1,293 tests pass. PR #140
79. GTM Agent — open-source go-to-market agent defaults to
80. re-agent — an agent built from first principles made Luna the default, with Sol in the catalog. PR #15
81. Everruns — Luna replaces GPT-5.6 Terra as the platform default. (PR) PR #3771
82. Spaire Claidor — retired GPT-5.6 Terra and Luna for GPT-6 Sol and Luna at half the cost. PR #239
83. Marinara Engine — multi-provider chat app adds both, with reasoning control. (PR) PR #6547
84. Geist — private local server for AI assistants adds Sol via OpenRouter. (PR) PR #378
85. News briefing generator — added an OpenAI adapter so Luna can stand in for Claude. PR #4
86. Siftdog — the company's Astro site now drafts blog content with Luna. PR #10
87. Legolizer (designs to LEGO models) — moved from gpt-5 to Sol; cost per build fell from about $0.20–0.30 to $0.13–0.14, and builds run 2–3× faster. PR #93
88. Conductor (AI MIDI music generation) — added both to its model registry. PR #112
89. Stagehand fork (browser automation) — adds Luna as a computer-use model for driving a browser. PR #7
90. TurenOS (security agent harness) — local-first security agent uses Sol and Luna as fallbacks. PR #83
91. FH Südwestfalen chat UI — a German university's OpenAI interface adds Luna as an option. (PR) PR #166
92. Teable — the no-code database replaced GPT-5.6 with Sol and Luna across AI Chat, App Builder, AI Fields and automations, and retired GPT-5.6 Terra. Teable
The reviewer role moved to Sol almost everywhere, often at high or extra-high reasoning. Luna shows up where the job is short and repetitive, like writing PR descriptions.
93. Ultralytics Actions — the YOLO team's GitHub automation made Luna the default for PR reviews and general requests, replacing GPT-5.6 Luna. Release v0.3.26
94. API Airforce SDK — CI code review runs on Sol at extra-high reasoning. PR #96
95. Shared Claude Code and Codex skills — reviewer skills moved from GPT-5.6 Sol to GPT-6 Sol at extra-high reasoning. PR #85
96. gdt (agent supervisor) — reviewer moved from GPT-5.6 Luna up to GPT-6 Sol. PR #50
97. Agentic coding project template — default reviewer moved to Sol on OpenRouter at the same $2/$10 price. PR #43
98. OpenClaw agent skills — auto-review defaults to Sol at high reasoning and retries on Luna if Sol is unavailable. (PR) PR #273
99. reVISit user-study framework — its adversarial-review subagent switches to Sol. (PR) PR #1479
100. b-open prompts — default coding worker moves from Grok 4.6 to Sol at extra-high reasoning. (PR) PR #96
101. Gradient Labs go-openai fork — added a
102. Payments adapter migration demo — a new repo where a Sol agent migrates a Python checkout service from Payments Adapter v1 to v2, for a documented $0.71 per run. Repo
103. StasisLang — an experimental Rust game-dev language defaults to Sol, and its "Gauntlet" roles run on Luna and escalate to Sol when needed. PR #823
104. AI project starter kit — moved its Luna and Sol agents from GPT-5.6 to GPT-6 across 10 files. PR #120
105. Wopee.io vibe-testing workshop — paid Copilot and gateway users switched from Claude Haiku 4.5 to Luna at a tenth of the price. PR #68
106. Checking another model's code review — a Hacker News user had Sol check 20 flags raised by another model and found about 75% were hallucinated. HN
107. Finding a real bug — in a side-by-side test, GPT-6 Sol found a cascade-delete bug, confirmed on PostgreSQL, that GPT-5.6 Sol missed. Codex issue #47787
108. tau (agent config) — subagents, review guidance, history digests and transcription spelling hints moved to Sol and Luna; 121 tests pass across six suites. PR #505
If you build on a framework, check it has shipped support before you switch: most of the big ones did so in the first two days.
109. Pydantic AI — model IDs plus reasoning and tool support, merged on launch day. PR #8635
110. Vercel AI SDK — model IDs for both, in v6 too. (PR) PR #21305
111. Mastra — listed as
112. Simon Willison's llm CLI — both model IDs in llm 0.36, per gptme's merged PR that depends on it. Issue #1702
113. chatsnack — Python chat library registers both as reasoning models; 114 tests pass. (PR) PR #89
114. Microsoft Amplifier OpenAI provider — model metadata, reasoning, cache handling and cost calculation. PR #113
115. promptfoo — the eval and red-teaming framework added both to its Chat, Responses and Codex catalogs, with cost tiers. PR #11043
116. models.dev — the open model database that many coding tools read from added metadata, pricing and Azure entries. PR #7781
117. Prime Intellect agent catalog — adds Luna via Azure. (PR) PR #12
If your cost dashboard doesn't know the new prices, it will report GPT-6 usage wrong. One RAG project found it had been under-counting costs about 5.3× (#53).
118. Langfuse — default pricing added so Sol and Luna costs are tracked automatically. PR #17802
119. aiusage — hard-codes Luna and Sol prices until LiteLLM's price list syncs. (PR) PR #71
120. Boxdeck (SSH dev-box cockpit) — usage pricing for Sol, Luna and Claude Opus 5.5. PR #44
121. endoflife.date — the product-lifecycle tracker lists both models as released September 22, 2026. (PR) PR #11114
These are tests people ran on their own tasks, not OpenAI's launch charts. Read them together: Sol usually lands close to Claude Opus 5.5 for a quarter to a third of the cost, and Luna is far cheaper but slips on harder tasks.
122. Cognition (Devin) — on its FrontierCode benchmark, Sol matched GPT-5.6 Sol at 61% lower cost per task, and Luna became the cheapest model on its leaderboard at under $0.10 per task. Devin blog
123. Sonar — on 4,444 Java tasks Sol passed 83.09% vs Astra's 85.85%, writing 27% less code with 22% fewer issues, but vulnerability density rose 60%. Sonar
124. Braintrust — on math tutoring, Luna scored 78.0% at about 2% of Opus 5.5's cost; Sol scored 84.1% on writing quality at about 25% of Opus's cost and was the fastest tested. Braintrust
125. Artificial Analysis — cost to run its Intelligence Index fell to $1.06 per task for Sol and $0.07 for Luna, and hallucination rates dropped from 92% to 60% (Sol) and 93% to 77% (Luna). Artificial Analysis
126. Vals AI — Sol ranks #8 on the Vals Index at 62.57%, but scored only 1.67% on Harvey's legal agent benchmark. Vals AI
127. Roboflow Playground (vision) — Sol averaged 82.3% across six vision tasks (#10 of 60) and Luna 77.2% (#18) at about $0.0004 per sample. Sol · Luna
128. Roboflow VLM exam (object detection) — mAP@50 of 73.6–75.2 for Sol and 56.8–64.1 for Luna, vs 82.1–83.6 for Astra. PR #50
129. ARC Prize — Luna scored 86.7% on ARC-AGI v1, 59.3% on v2 and under 1% on v3. ARC Prize
130. Vercel Next.js agent evals — Sol passed 30 of 31 at $0.24 per eval, the same as Opus 5.5; Luna passed 24 of 31 (27 with doc tweaks) at $0.011. PR #127
131. githits-cli agent evals — moving from GPT-5.6 Luna halved cost from $0.477 to $0.241 and raised valid runs from 49 to 52 of 52, about 10% slower. PR #415
132. Uzbek-language benchmark — Luna scored 0.8087 on 1,861 questions, third of 72 models. PR #14
133. ReuseBeacon — a 72-run code-reuse benchmark on Luna at high reasoning; each arm passed 21 of 24 and every score reproduced. Release
134. Do models memorize earnings results? — a study of 12 models; Sol scored AUC 0.691 and Luna 0.573 at "remembering" 2023 earnings beats from a name and date. dev.to
135. DataCamp — building a gravity-flipping Tetris clone cost $0.26 on Sol vs $0.87 on Opus 5.5, and Sol scored higher, 5.0 vs 4.3. DataCamp
136. The New Stack — CI triage, a 3,664-line incident log and a spec task, five runs each: Sol was perfect on 12 of 15 vs 15 of 15 for Opus, but cost $2.68 vs $16.72. The New Stack
137. Analytics Vidhya — Sol found both planted SQL bugs plus an extra one; Luna got the business math right and correctly found no meeting slot existed. Analytics Vidhya
138. MindStudio — a landing page cost $5.89 on Sol vs $18.32 on Opus 5.5, but Sol returned a wrong date and Opus's output looked more polished. MindStudio
139. Simon Willison's pelicans — SVG pelicans at every reasoning level cost 0.89–18.23¢ on Sol and 0.07–0.91¢ on Luna. simonwillison.net
140. Lenny's Newsletter — in Claire Vo's blind taste test, Sol was best for clear writing and readable PRDs, and Opus 5.5 for long-running agents. Lenny's Newsletter
141. The Neuron — livestreamed game and 3D builds; Sol was cheaper for routine work, Opus better at long autonomous runs. The Neuron
142. Microsoft 365 Copilot test — Sol built a 3-year SaaS financial model in 5–7 minutes vs over an hour for Opus 5.5, for about $7 vs $10 in credits. Substack
143. Qiita — Luna in Copilot updated every VS Code extension in 16 steps and 1m41s for about $0.014. Qiita
144. Zenn — Sol polished bilingual study notes for 33 real-estate exam questions; the author saw "no major advance" over GPT-5.6 Sol. Zenn
145. "Codexometer" on the OpenAI forum — on three coding tasks Luna ranked first and Sol second; one scheduler task cost about $0.0023 on Luna vs $0.0575 on Sol. OpenAI forum
146. Sol reasoning-effort guide — a repo on which Sol reasoning setting to use for cited research. Repo
Ten cases where GPT-6 didn't live up to the launch post. Check these before you switch a production workload.
147. Luna's Kubernetes files crashed on deploy — on a real k3s cluster, Luna's manifests passed the linter but crash-looped on deploy (0 of 3 worked). Sol passed 9 of 9 DevOps tasks for $0.0841, 69% cheaper than Opus 5.5. ComputingForGeeks
148. Cache-heavy work can cost more — one developer found a cached coding task cost $0.072 on Luna vs $0.019 on Xiaomi's MiMo 2.6, and estimated $184 a month vs about $40 today. HN
149. Sol "kept rabbit holing" in Go — another developer said Gemini 3.8 Flash caught architecture problems that Sol missed. HN
150. More bugs than GPT-5.6 Sol — on cubic's Bug Bench, Sol's relative bug rate was 1.38, worse than GPT-5.6 Sol's 0.72. cubic
151. Lower pass rate in Codex, so far — MarginLab's daily tracker shows Sol at 64% over 7 days vs 83.4% for the GPT-5.6 Sol baseline, though the new baseline is still being collected. MarginLab
152. Slower Codex tasks — users reported tasks taking about 50 minutes vs 20 on GPT-5.6 Sol, later 25–30; it may not be the model's fault. Codex issue #47656
153. Cline on Bedrock — bare Bedrock model IDs return errors; a fix is proposed. Cline issue #14468
154. LangChain on Bedrock — GPT-6 calls fail with 400 errors on tool calls and temperature. LangChain issue #40810
155. Not every ChatGPT surface has it — one project found chat requests for Sol or Luna were silently served by GPT-5.6, because the models are limited to Work and Codex. PR #76
156. No GPT-6 in Cursor — OpenAI announced on August 31 that it is ending its Cursor partnership after SpaceX acquired Cursor, so new models like GPT-6 aren't offered there; your own API key or OpenAI's Codex extension are the workarounds. Tom's Guide
If you shipped something on GPT-6 Sol or Luna, publish it on HackerNoon and it may make the next version of this list.
OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026. Sol costs $2 per million input tokens and $10 per million output tokens, and it is aimed at coding and agent work. Luna costs $0.10 in and $0.50 out, and it is aimed at high-volume extraction, summarizing, classification and routing. Both are about half the price of the GPT-5.6 models they replace, and cached input is 90% off (OpenAI, VentureBeat).
Every example below links to something you can check: a merged pull request, a release, a platform's docs, or a write-up with numbers. Entries marked (PR) are pull requests that were not yet merged when this was written. The last section collects what went wrong, because a list of only wins is not much use when you're deciding what to build.
Sol or Luna? What the first week shows
The pattern across these projects: Luna goes on work that is high-volume and has a clear right answer, and Sol goes on coding, review and anything that needs judgment.
- Extraction from invoices, forms and posts → Luna. 93% of invoice fields correct vs 82% on gpt-4o, at about a fifth of the cost (#49).
- Classification and tagging → Luna. Same accuracy as GPT-5.6 Luna on whisky photos, 31% cheaper (#61).
- Chatbots and default app models → Luna, with Sol as the upgrade. One chatbot builder reports Luna is "~7-9x cheaper" than its old default (#71).
- Code review in CI → Sol, often at high reasoning. Many repos moved their reviewer from GPT-5.6 Sol to GPT-6 Sol (#93–#100).
- Infrastructure code (Kubernetes, Terraform) → Sol. Luna's Kubernetes files passed linting but crashed on deploy; Sol passed 9 of 9 (#147).
- Cache-heavy long coding sessions → test before switching. One developer estimated $184 a month on Luna vs about $40 on a cheaper rival (#148).
Where you can call it: APIs, clouds and gateways (1–26)
You don't need an OpenAI account to use GPT-6. Within a week it was available through every major cloud except Google, and through most model gateways.
1. OpenAI API —
gpt-6-sol and gpt-6-luna in the Responses and Chat Completions APIs, both with about a 1M-token context window. Model page2. ChatGPT Work and desktop app — paid plans get both models in Work and Codex; free users get Luna, but only in the desktop app. Notebookcheck
3. Microsoft Foundry (Azure) — Astra, Sol and Luna in 28 regions, with Luna pitched for extraction, routing and classification. Azure blog
4. Amazon Bedrock — Sol for coding and data analysis, Luna for extraction and Q&A at scale, with prompt caching and optional zero data retention. AWS blog
5. Snowflake Cortex AI — call Sol or Luna straight from SQL through Cortex AI Functions (public preview). Snowflake
6. Databricks Foundation Model APIs — both models listed as supported. Databricks docs
7. Microsoft 365 Copilot and Copilot Studio — Sol (not Luna) in Word, Excel, PowerPoint and Copilot Studio, next to Claude Opus 5.5. Windows Report
8. OpenRouter — both models, routed across OpenAI, Azure and Bedrock. In week one Codex ran 166B Sol tokens through it, Hermes Agent 82.5B and pi 53.5B. Sol · Luna
9. Vercel AI Gateway — same-day support; Sol for complex agent work, Luna as the cheaper option. Changelog
10. Cloudflare AI — both models; Sol goes to $4/$15 for long-context requests. Sol · Luna
11. LiteLLM — works from launch day on v1.101.0+, with pricing in its cost map. LiteLLM blog
12. Requesty — Standard and Flex tiers, with Luna pitched for structured extraction. Requesty
13. TrueFoundry AI Gateway — suggests routing each request to Sol or Luna by complexity and claims 50–70% savings. TrueFoundry
14. Portkey — pricing and capability entries in its open model catalog. PR #947
15. CometAPI — resells Sol input at $1.60 per million, 20% under OpenAI's price. CometAPI
16. EvoLink — Luna through its Chat Completions and Responses endpoints. EvoLink
17. NEAR AI Cloud — fixed tool calling so both models work with reasoning turned off. PR #1121
18. Pollinations — open generative-AI API added both as paid-only models. PR #15312
19. OpenGradient TEE Gateway — both models served from a hardware-secured enclave. PR #171
20. ClawRouter — multi-provider router added endpoints, limits and pricing for both. PR #247
21. Coder's xum — routing layer adds both, keeping old model IDs working. (PR) PR #4348
22. Windmill Luna Gateway — an OpenAI-compatible gateway for Windmill moved its default from GPT-5.6 Luna to GPT-6 Luna. PR #13
23. Felo AI — Sol in Felo Search and the Felo API, for batch summarizing and extraction. Felo
24. Emergent — Sol available through its single "Universal Key" for app building. Emergent
25. Puter.js — free access to both models from the browser with no API key; Puter's cost model puts an 80-turn coding session at $2.32 on Sol and $0.116 on Luna, with caching saving more than the rate cut. Puter
26. Make — the no-code automation platform added both to its AI Toolkit, OpenAI app and AI Agents, suggesting Luna for categorizing and routing support requests and Sol for turning bug reports into fix drafts. Make community
In your editor and coding agent (27–48)
The fastest adoption was in coding tools. Most made Sol the default for writing code and kept Luna for small, fast tasks.
27. OpenAI Codex — Sol and Luna added to the model catalog, with migration paths from GPT-5.x. Simon Willison made Sol his Codex default. PR #47332
28. GitHub Copilot — Sol and Luna in VS Code, Visual Studio, JetBrains, Xcode, Eclipse, the CLI and github.com. Changelog
29. Zed — both models for users who bring their own OpenAI key. PR #64628
30. Warp — each model at five reasoning levels, low to max. Warp docs
31. Factory (Droid CLI) — Luna is billed at 0.04× a standard request and Sol at 0.8×, so Luna is the cheap default for routine edits. Factory docs
32. Kilo Code — both models; Kilo estimates about $0.18 to have Sol analyze a 10,000-line codebase. Sol · Luna
33. OpenCode — both models through ChatGPT/Codex sign-in. PR #50704
34. Pi (pi.dev) — terminal coding agent offers Sol through Copilot, Azure and OpenAI. Pi
35. Codex plugin for Claude Code — OpenAI's plugin made Sol the default for new tasks. PR #784
36. gptme — terminal agent added both to its model registry. PR #3914
37. maki — Lua-scriptable coding agent added both for ChatGPT Coding Plan users (a fork, makima, did the same). PR #1088 · makima #145
38. Overdeck — multi-agent orchestrator for coding assistants registered both in every picker. PR #4234
39. Phoenix IDE — both models on direct OpenAI and Codex routes. PR #801
40. microcodex — ultra-light terminal agent switched its default from GPT-5.6 Sol to GPT-6 Sol. PR #23
41. ChatGPT for VS Code (community fork) — added both, with Sol as the default. Release
42. Background Agents — open-source background coding agents added both to their Slack, Linear and GitHub bots. PR #2028
43. opencodex — ranks GPT-6 ahead of GPT-5.6 Sol across providers. (PR) PR #5584
44. makeitworkcloud Helm charts — moved 9 production OpenCode agents from GPT-5.6 Terra to GPT-6 Sol. PR #114
45. A developer's dotfiles — Pi coding-agent routes moved to GPT-6, with GPT-5.6 Terra kept as a fallback. PR #31
46. Paperclip — open-source agent manager adds Sol to its Codex adapter, with reasoning from low to ultra. (PR) PR #13830
47. GItS — bumped its Codex CLI so Sol is the default. PR #655
48. OpenClaw Code Agent — allows Luna and prices both in its cost estimator. (PR) PR #479
Extraction, RAG and document processing (49–58)
This is where Luna has the clearest numbers so far: cheaper than the model it replaced, and in several cases more accurate too.
49. Facturear (invoice extraction) — on 22 real invoices, Luna got 93.2% of fields right vs 82.5% for gpt-4o. Tax IDs were right on 21 of 22 vs 10 of 22, median time fell from 36–44s to 15.5s, and cost fell from $0.021–0.034 to about $0.0044 per invoice. PR #37
50. EICR (live extraction backend) — moving from GPT-5.6 Luna cut median latency from 4,494 to 3,354 ms and cost by 55%, with all 13 test fixtures matching. PR #221
51. reflection (claim extraction) — 56 of 56 outputs valid and 58% cheaper than GPT-5.6 Luna. Precision rose from 57% to 74% while recall fell from 85% to 73%. PR #28
52. which-glp (post extraction) — Luna was at least as accurate as gpt-5-nano, which misread "8 lbs" once, at about 1.4× nano's cost. PR #214
53. Catholic theology search (RAG) — swapping gpt-5.4-mini for Luna raised answer quality from 4.19 to 4.67 out of 5 and cut unsupported claims from 16.7% to 1.7%, at about 12% lower cost. PR #114
54. Voicebook — a local-first document reader with text-to-speech made Luna its default, replacing the GPT-5.6 models. PR #159
55. The Plow Times — a research-paper generator that delivers sourced PDFs moved from Claude Opus 5 to GPT-6 Sol via OpenRouter, using the 1M-token context. PR #132
56. Sure (personal finance) — adds Sol for chat, vision and PDF processing of statements. (PR) PR #3771
57. reviewer2 (academic paper reviewer) — fixed failing GPT-6 requests and added GPT-6 pricing, moving its docs off gpt-5-mini. PR #3
58. PaperPilot (Zotero research tool) — Sol and Luna added to its Codex model picker for working through papers. PR #76
Classification, tagging and routing (59–66)
Put something in a bucket, pick a label, decide where a request goes: this is the work Luna was priced for.
59. Microsoft openaivec — Microsoft's SDK for running OpenAI models over Spark and DuckDB made Luna its new default, replacing gpt-4.1-mini for classification, extraction and translation. Release v2.7.0
60. itemize (Monarch Money transactions) — Luna categorizes retail transactions, replacing GPT-5.6 Luna at $0.10/$0.50 vs $0.20/$1.20 per million tokens. PR #97
61. Peated (whisky-bottle photos) — David Cramer's classifier matched GPT-5.6 Luna at 79 of 105 correct, and cost fell 31% from $0.32 to $0.22, though runs got slower, from 22.5s to 31.4s. PR #1311
62. phototag (Lightroom photo tagging) — default vision model moves from gpt-4o-mini to Luna. (PR) PR #1
63. Telegram bot — Luna decides whether the bot should reply and handles web search, replacing GPT-5.6 Luna. PR #75
64. jev-issue-router — recommends model and reasoning settings for each GitHub issue; added Sol and Luna in place of GPT-5.6 Terra. PR #5
65. Bühler Group internal chat — the industrial company's chat app makes Sol the default and drops to Luna when spending caps are hit. (PR) PR #182
66. kipp workflows — replaced GPT-5.6 Luna with the
gpt-luna-latest alias, which now points to GPT-6 Luna. Issue #93Chatbots, assistants and product agents (67–92)
The most common move in week one: make Luna the default model in a product, and keep Sol for the users or tasks that need more.
67. Dust — the enterprise agent platform made Luna the default for its built-in agents, citing half the input price and 2.4× cheaper output than GPT-5.6 Luna. PR #33031 · docs
68. Perplexity — Sol is the default for the "Light" effort setting, including its Computer agent. Crypto Briefing
69. Datasette Agent — Simon Willison moved the agent.datasette.io demo to Luna and calls it "fast and competent" at SQL queries and HTML/JS apps. simonwillison.net
70. Hermes Agent (Nous Research) — swaps its GPT-5.6 tiers for GPT-6 and is the #2 app by Sol tokens on OpenRouter. (PR) PR #119410
71. chatniuexa (chatbot builder) — Luna replaces gpt-5.4-mini as the default, and bots that were on flagship models move to Sol; the team says Luna is "~7-9x cheaper". PR #2
72. VYNN AI stock analyst — Luna replaces gpt-5.4-mini for chat and gpt-4o-mini for analysis; a full analysis ran with no API errors. PR #22
73. TradingAgents — the multi-agent trading framework uses Sol for analysis, a scheduled writer/judge, YouTube synthesis and portfolio decisions. PR #288
74. TollChat (Northern Virginia commuter finance agent) — the app, its simulated users and its judges all moved to Luna; a 600-trial baseline passed 68.97% of cases for $1.72 total. PR #590
75. In-game character agents (CS3216 project) — p90 latency fell from 3.91s to 2.43s and cost per call from $0.00094 to $0.00041, with 30 of 30 valid outputs and 5 of 5 safe replies to hostile prompts. PR #138
76. Cadencia-AI (goal planner) — moved from DeepSeek to Luna in production; a goal run went from 65,472 to 23,584 micro-dollars. PR #30
77. Open Science Assistant — default set to Luna on Bedrock; three community assistants answered live with citations. PR #529
78. Stately Agent — David Khourshid's state-machine agent library made Luna its demo default, with Sol for one script; 1,293 tests pass. PR #140
79. GTM Agent — open-source go-to-market agent defaults to
gpt-6-luna-fast at high reasoning. PR #7780. re-agent — an agent built from first principles made Luna the default, with Sol in the catalog. PR #15
81. Everruns — Luna replaces GPT-5.6 Terra as the platform default. (PR) PR #3771
82. Spaire Claidor — retired GPT-5.6 Terra and Luna for GPT-6 Sol and Luna at half the cost. PR #239
83. Marinara Engine — multi-provider chat app adds both, with reasoning control. (PR) PR #6547
84. Geist — private local server for AI assistants adds Sol via OpenRouter. (PR) PR #378
85. News briefing generator — added an OpenAI adapter so Luna can stand in for Claude. PR #4
86. Siftdog — the company's Astro site now drafts blog content with Luna. PR #10
87. Legolizer (designs to LEGO models) — moved from gpt-5 to Sol; cost per build fell from about $0.20–0.30 to $0.13–0.14, and builds run 2–3× faster. PR #93
88. Conductor (AI MIDI music generation) — added both to its model registry. PR #112
89. Stagehand fork (browser automation) — adds Luna as a computer-use model for driving a browser. PR #7
90. TurenOS (security agent harness) — local-first security agent uses Sol and Luna as fallbacks. PR #83
91. FH Südwestfalen chat UI — a German university's OpenAI interface adds Luna as an option. (PR) PR #166
92. Teable — the no-code database replaced GPT-5.6 with Sol and Luna across AI Chat, App Builder, AI Fields and automations, and retired GPT-5.6 Terra. Teable
Code review, CI and dev automation (93–108)
The reviewer role moved to Sol almost everywhere, often at high or extra-high reasoning. Luna shows up where the job is short and repetitive, like writing PR descriptions.
93. Ultralytics Actions — the YOLO team's GitHub automation made Luna the default for PR reviews and general requests, replacing GPT-5.6 Luna. Release v0.3.26
94. API Airforce SDK — CI code review runs on Sol at extra-high reasoning. PR #96
95. Shared Claude Code and Codex skills — reviewer skills moved from GPT-5.6 Sol to GPT-6 Sol at extra-high reasoning. PR #85
96. gdt (agent supervisor) — reviewer moved from GPT-5.6 Luna up to GPT-6 Sol. PR #50
97. Agentic coding project template — default reviewer moved to Sol on OpenRouter at the same $2/$10 price. PR #43
98. OpenClaw agent skills — auto-review defaults to Sol at high reasoning and retries on Luna if Sol is unavailable. (PR) PR #273
99. reVISit user-study framework — its adversarial-review subagent switches to Sol. (PR) PR #1479
100. b-open prompts — default coding worker moves from Grok 4.6 to Sol at extra-high reasoning. (PR) PR #96
101. Gradient Labs go-openai fork — added a
GPT6Luna constant so the team can write PR descriptions with Luna. PR #45102. Payments adapter migration demo — a new repo where a Sol agent migrates a Python checkout service from Payments Adapter v1 to v2, for a documented $0.71 per run. Repo
103. StasisLang — an experimental Rust game-dev language defaults to Sol, and its "Gauntlet" roles run on Luna and escalate to Sol when needed. PR #823
104. AI project starter kit — moved its Luna and Sol agents from GPT-5.6 to GPT-6 across 10 files. PR #120
105. Wopee.io vibe-testing workshop — paid Copilot and gateway users switched from Claude Haiku 4.5 to Luna at a tenth of the price. PR #68
106. Checking another model's code review — a Hacker News user had Sol check 20 flags raised by another model and found about 75% were hallucinated. HN
107. Finding a real bug — in a side-by-side test, GPT-6 Sol found a cascade-delete bug, confirmed on PostgreSQL, that GPT-5.6 Sol missed. Codex issue #47787
108. tau (agent config) — subagents, review guidance, history digests and transcription spelling hints moved to Sol and Luna; 121 tests pass across six suites. PR #505
Agent frameworks, SDKs and libraries (109–117)
If you build on a framework, check it has shipped support before you switch: most of the big ones did so in the first two days.
109. Pydantic AI — model IDs plus reasoning and tool support, merged on launch day. PR #8635
110. Vercel AI SDK — model IDs for both, in v6 too. (PR) PR #21305
111. Mastra — listed as
openai/gpt-6-sol and openai/gpt-6-luna. Mastra docs112. Simon Willison's llm CLI — both model IDs in llm 0.36, per gptme's merged PR that depends on it. Issue #1702
113. chatsnack — Python chat library registers both as reasoning models; 114 tests pass. (PR) PR #89
114. Microsoft Amplifier OpenAI provider — model metadata, reasoning, cache handling and cost calculation. PR #113
115. promptfoo — the eval and red-teaming framework added both to its Chat, Responses and Codex catalogs, with cost tiers. PR #11043
116. models.dev — the open model database that many coding tools read from added metadata, pricing and Azure entries. PR #7781
117. Prime Intellect agent catalog — adds Luna via Azure. (PR) PR #12
Cost tracking and observability (118–121)
If your cost dashboard doesn't know the new prices, it will report GPT-6 usage wrong. One RAG project found it had been under-counting costs about 5.3× (#53).
118. Langfuse — default pricing added so Sol and Luna costs are tracked automatically. PR #17802
119. aiusage — hard-codes Luna and Sol prices until LiteLLM's price list syncs. (PR) PR #71
120. Boxdeck (SSH dev-box cockpit) — usage pricing for Sol, Luna and Claude Opus 5.5. PR #44
121. endoflife.date — the product-lifecycle tracker lists both models as released September 22, 2026. (PR) PR #11114
Benchmarks and hands-on tests (122–146)
These are tests people ran on their own tasks, not OpenAI's launch charts. Read them together: Sol usually lands close to Claude Opus 5.5 for a quarter to a third of the cost, and Luna is far cheaper but slips on harder tasks.
122. Cognition (Devin) — on its FrontierCode benchmark, Sol matched GPT-5.6 Sol at 61% lower cost per task, and Luna became the cheapest model on its leaderboard at under $0.10 per task. Devin blog
123. Sonar — on 4,444 Java tasks Sol passed 83.09% vs Astra's 85.85%, writing 27% less code with 22% fewer issues, but vulnerability density rose 60%. Sonar
124. Braintrust — on math tutoring, Luna scored 78.0% at about 2% of Opus 5.5's cost; Sol scored 84.1% on writing quality at about 25% of Opus's cost and was the fastest tested. Braintrust
125. Artificial Analysis — cost to run its Intelligence Index fell to $1.06 per task for Sol and $0.07 for Luna, and hallucination rates dropped from 92% to 60% (Sol) and 93% to 77% (Luna). Artificial Analysis
126. Vals AI — Sol ranks #8 on the Vals Index at 62.57%, but scored only 1.67% on Harvey's legal agent benchmark. Vals AI
127. Roboflow Playground (vision) — Sol averaged 82.3% across six vision tasks (#10 of 60) and Luna 77.2% (#18) at about $0.0004 per sample. Sol · Luna
128. Roboflow VLM exam (object detection) — mAP@50 of 73.6–75.2 for Sol and 56.8–64.1 for Luna, vs 82.1–83.6 for Astra. PR #50
129. ARC Prize — Luna scored 86.7% on ARC-AGI v1, 59.3% on v2 and under 1% on v3. ARC Prize
130. Vercel Next.js agent evals — Sol passed 30 of 31 at $0.24 per eval, the same as Opus 5.5; Luna passed 24 of 31 (27 with doc tweaks) at $0.011. PR #127
131. githits-cli agent evals — moving from GPT-5.6 Luna halved cost from $0.477 to $0.241 and raised valid runs from 49 to 52 of 52, about 10% slower. PR #415
132. Uzbek-language benchmark — Luna scored 0.8087 on 1,861 questions, third of 72 models. PR #14
133. ReuseBeacon — a 72-run code-reuse benchmark on Luna at high reasoning; each arm passed 21 of 24 and every score reproduced. Release
134. Do models memorize earnings results? — a study of 12 models; Sol scored AUC 0.691 and Luna 0.573 at "remembering" 2023 earnings beats from a name and date. dev.to
135. DataCamp — building a gravity-flipping Tetris clone cost $0.26 on Sol vs $0.87 on Opus 5.5, and Sol scored higher, 5.0 vs 4.3. DataCamp
136. The New Stack — CI triage, a 3,664-line incident log and a spec task, five runs each: Sol was perfect on 12 of 15 vs 15 of 15 for Opus, but cost $2.68 vs $16.72. The New Stack
137. Analytics Vidhya — Sol found both planted SQL bugs plus an extra one; Luna got the business math right and correctly found no meeting slot existed. Analytics Vidhya
138. MindStudio — a landing page cost $5.89 on Sol vs $18.32 on Opus 5.5, but Sol returned a wrong date and Opus's output looked more polished. MindStudio
139. Simon Willison's pelicans — SVG pelicans at every reasoning level cost 0.89–18.23¢ on Sol and 0.07–0.91¢ on Luna. simonwillison.net
140. Lenny's Newsletter — in Claire Vo's blind taste test, Sol was best for clear writing and readable PRDs, and Opus 5.5 for long-running agents. Lenny's Newsletter
141. The Neuron — livestreamed game and 3D builds; Sol was cheaper for routine work, Opus better at long autonomous runs. The Neuron
142. Microsoft 365 Copilot test — Sol built a 3-year SaaS financial model in 5–7 minutes vs over an hour for Opus 5.5, for about $7 vs $10 in credits. Substack
143. Qiita — Luna in Copilot updated every VS Code extension in 16 steps and 1m41s for about $0.014. Qiita
144. Zenn — Sol polished bilingual study notes for 33 real-estate exam questions; the author saw "no major advance" over GPT-5.6 Sol. Zenn
145. "Codexometer" on the OpenAI forum — on three coding tasks Luna ranked first and Sol second; one scheduler task cost about $0.0023 on Luna vs $0.0575 on Sol. OpenAI forum
146. Sol reasoning-effort guide — a repo on which Sol reasoning setting to use for cited research. Repo
Where it fell short (147–156)
Ten cases where GPT-6 didn't live up to the launch post. Check these before you switch a production workload.
147. Luna's Kubernetes files crashed on deploy — on a real k3s cluster, Luna's manifests passed the linter but crash-looped on deploy (0 of 3 worked). Sol passed 9 of 9 DevOps tasks for $0.0841, 69% cheaper than Opus 5.5. ComputingForGeeks
148. Cache-heavy work can cost more — one developer found a cached coding task cost $0.072 on Luna vs $0.019 on Xiaomi's MiMo 2.6, and estimated $184 a month vs about $40 today. HN
149. Sol "kept rabbit holing" in Go — another developer said Gemini 3.8 Flash caught architecture problems that Sol missed. HN
150. More bugs than GPT-5.6 Sol — on cubic's Bug Bench, Sol's relative bug rate was 1.38, worse than GPT-5.6 Sol's 0.72. cubic
151. Lower pass rate in Codex, so far — MarginLab's daily tracker shows Sol at 64% over 7 days vs 83.4% for the GPT-5.6 Sol baseline, though the new baseline is still being collected. MarginLab
152. Slower Codex tasks — users reported tasks taking about 50 minutes vs 20 on GPT-5.6 Sol, later 25–30; it may not be the model's fault. Codex issue #47656
153. Cline on Bedrock — bare Bedrock model IDs return errors; a fix is proposed. Cline issue #14468
154. LangChain on Bedrock — GPT-6 calls fail with 400 errors on tool calls and temperature. LangChain issue #40810
155. Not every ChatGPT surface has it — one project found chat requests for Sol or Luna were silently served by GPT-5.6, because the models are limited to Work and Codex. PR #76
156. No GPT-6 in Cursor — OpenAI announced on August 31 that it is ending its Cursor partnership after SpaceX acquired Cursor, so new models like GPT-6 aren't offered there; your own API key or OpenAI's Codex extension are the workarounds. Tom's Guide
What the first week tells you
- Start with Luna for anything that has a right answer. Extraction, tagging and routing projects cut costs 30–80% and usually kept or improved accuracy.
- Use Sol where the output gets executed or reviewed. Code review, infrastructure code and agent planning moved to Sol, not Luna.
- Measure your own workload before you switch. Caching, latency and your framework's support vary, and several teams saw slower runs.
- Update your cost tracking first. Otherwise you can't tell whether the switch saved anything.
If you shipped something on GPT-6 Sol or Luna, publish it on HackerNoon and it may make the next version of this list.