GPT-6.2 vs Claude side by side
API list prices per million tokens at standard speed. OpenAI's prices apply up to 272K input tokens; above that, input and cache rates double and output is 1.5x for the whole request.
| GPT-6.2 Astra | GPT-6.2 Sol | Claude Fable 5.2 | Claude Opus 5.5 | |
|---|---|---|---|---|
| Maker | OpenAI | OpenAI | Anthropic | Anthropic |
| API model ID | gpt-6.2-astra | gpt-6.2-sol | claude-fable-5-2 | claude-opus-5-5 |
| Released | November 17, 2026 | November 17, 2026 | November 9, 2026 | September 22, 2026 |
| Price (input / output) | $8 / $40 | $2 / $10 | $8 / $40 | $4 / $20 |
| Cached input / cache reads | $0.40 | $0.10 | $0.20 | $0.20 |
| Batch | 50% off | 50% off | 50% off | 50% off |
| Context window | 1.05M tokens (922K max input) | 1.05M tokens (922K max input) | 1M tokens | 1M tokens |
| Max output | 128K | 128K | 128K | 128K |
| Knowledge cutoff | May 18, 2026 | June 1, 2026 | June 2026 | June 2026 |
| Input / output | Text and image in, text out | Text and image in, text out | Text and image in, text out | Text and image in, text out |
| Effort | low, medium, high, xhigh, max; none is rejected; OpenAI suggests starting at medium | Same levels as Astra | Adaptive thinking, always on; low to max, default high | Adaptive thinking, always on; default medium |
| Faster modes | Fast (2x price), Ultrafast (3x price through February 28, 2027, then 6x) | Fast (2x price), Ultrafast (3x price through February 28, 2027, then 6x) | Fast mode, up to 2.5x faster, at $16 / $80 | Fast mode, up to 2.5x faster, at $8 / $40 |
| Coding tool | Codex app and CLI (/model); GitHub Copilot | Codex app and CLI (/model); GitHub Copilot | Claude Code, picked with /model (v2.1.321 or later) | Claude Code's default model (v2.1.280 or later) |
| Plans | ChatGPT Plus, Pro, Enterprise and Edu, in ChatGPT Work and Codex | ChatGPT Plus, Pro, Enterprise and Edu, in ChatGPT Work and Codex | Claude Pro, Max, Ultra, Team and Enterprise | Claude Pro, Max, Ultra, Team and Enterprise |
| Also on | OpenAI API, GitHub Copilot, Microsoft Foundry | OpenAI API, GitHub Copilot | Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry | Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry |
| What the maker says it's for | “Your hardest projects”, such as tracing bugs across services | “Everyday work”, everyday coding and agent workflows | Demanding reasoning and long-horizon agentic work, or when Opus 5.5 at higher effort falls short | Long-running agentic coding and knowledge work; Anthropic's recommended starting model |
Benchmarks: what each vendor reported
Each number is labelled with who reported it. OpenAI and Anthropic both publish FrontierCode 2.0, Terminal-Bench-Science and AutomationBench, but each ran its own models with its own harness, effort setting and safeguards, so a row is two separate reports side by side, not one head-to-head run. Neither vendor has published GPT-6.2 and Claude in the same setup.
OpenAI's GPT-6.2 numbers come from its launch post (as reported by VentureBeat, which notes each model's best plotted score is used, so effort may differ). Anthropic's numbers come from its Fable page, at the highest available effort with safeguards on.
| Benchmark | GPT-6.2 Astra | GPT-6.2 Sol | Claude Fable 5.2 | Claude Opus 5.5 |
|---|---|---|---|---|
| FrontierCode 2.0 (agentic coding) | 44.2% (OpenAI) | 43.8% (OpenAI) | 49.1% at xhigh (Anthropic) | 45.5% (Anthropic) |
| Terminal-Bench-Science 0.1 | 77.1% (OpenAI) | 68.0% (OpenAI) | 74.6% at max, in Claude Code (Anthropic) | 58.7% (Anthropic) |
| AutomationBench (business workflows) | 52.4% (OpenAI) | 41.2% (OpenAI) | 51.4% (Anthropic) | 42.5% (Anthropic) |
| FrontierSWE v2 (agentic coding) | Not reported | Not reported | 69.9% (Anthropic) | 62.3% (Anthropic) |
| OSWorld 2.1 (computer use) | Not reported | Not reported | 89.3% (Anthropic) | 87.2% (Anthropic) |
| Agents' Last Exam v1.1 | 78.3% (OpenAI) | 73.0% (OpenAI) | 5.5 points below Astra in OpenAI's chart; that leaderboard excludes nine Fable 5.2 tasks that used a fallback model | Not reported |
| Harvey LAB-AA v1.1 (legal) | 15.0% (OpenAI) | 11.1% (OpenAI) | OpenAI says Astra more than doubles Fable 5.2's pass rate | Not reported |
| SWE-bench | Not reported | Not reported | Not reported | Not reported |
Reading the vendor numbers
- Terminal-Bench-Science: OpenAI's chart also shows a Fable 5.2 score, but OpenAI's footnote says that figure is Artificial Analysis's run in mini-swe-agent. Anthropic reported 74.6% at max reasoning in Claude Code. Use Anthropic's figure for Anthropic's model and OpenAI's for OpenAI's.
- FrontierCode 2.0: OpenAI says Sol came within 0.5 points of Astra at about one fifth of the cost, both at max effort. That's one benchmark, not a general cost ratio.
- Agents' Last Exam and Harvey LAB-AA are OpenAI's comparisons against Fable 5.2; Anthropic hasn't published numbers on them.
- FrontierSWE v2 and OSWorld 2.1 are Anthropic's; OpenAI didn't report GPT-6.2 on them.
Independent indexes
Rankings run by third parties rather than by OpenAI or Anthropic. Artificial Analysis and Epoch AI numbers are as reported by VentureBeat; AA-3D is the Artificial Analysis leaderboard OpenAI cites in its launch post. These reports didn't include Opus 5.5.
| Index | GPT-6.2 Astra | GPT-6.2 Sol | Claude Fable 5.2 |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 62, second place (Artificial Analysis) | - | First place, ahead of Astra (Artificial Analysis) |
| Epoch AI index | 174 (Epoch) | - | 174, tied with Astra (Epoch) |
| AA-3D (blinded human votes on 3D and browser builds) | 1,318 Elo (Artificial Analysis) | 1,259 Elo (Artificial Analysis) | 1,271 Elo (Artificial Analysis) |
| Artificial Analysis Coding Agent Index (Codex, max effort) | Flat against the previous version (Artificial Analysis) | Down 3 points, largely from a 14-point drop on SWE-Atlas-QnA-AA (Artificial Analysis) | - |
| One developer's same-prompt test (@EnvolDev on X, "Cathedral of Ice" prompt; not a benchmark) | $8.20, 40m 34s; the author judged its quality well below Fable 5.2 | - | $81.10, 1h 56m 52s |
GPT-6.2 Astra vs Claude Fable 5.2
These are the two flagships, and they cost the same per token: $8 per million input and $40 per million output. Both take a 1M-class context and write up to 128K tokens. On the independent indexes they're close: tied at 174 on Epoch's index, Fable 5.2 ahead on Artificial Analysis's Intelligence Index, Astra ahead on AA-3D.
- What they're for: OpenAI calls Astra its model for “your hardest projects”, such as tracing bugs across services. Anthropic positions Fable 5.2 for demanding reasoning and long-horizon agent work, and says it can run ambiguous engineering tasks for days.
- Caching: Fable 5.2's cache reads cost $0.20 per million, half of Astra's $0.40 cached input. On long agent sessions that reuse the same prefix, that difference adds up.
- Effort: Fable 5.2 defaults to
high. OpenAI suggests starting Astra atmedium. Both go up tomax, and higher effort usually means more tokens per task, so compare cost per finished task as well as list price. - Long inputs: above 272K input tokens, OpenAI doubles Astra's input price and raises output 1.5x for the whole request.
- Speed: Fable 5.2's fast mode is up to 2.5x faster at double the price. Astra has Fast (2x price) and Ultrafast (3x price until February 28, 2027) tiers; Ultrafast is US data residency only.
- Where you run it: Astra in Codex, ChatGPT Work and GitHub Copilot; Fable 5.2 in Claude Code, where you pick it with
/model. On Claude Pro and standard Team seats, Fable 5.2 can use up to 50% of your weekly limits. - Safeguards: both restrict some cybersecurity work. OpenAI says its tighter cyber safeguards also block some legitimate security research; Anthropic routes flagged cyber queries on Fable 5.2 to Claude Opus 4.8.
- Data retention: Fable models keep data for 30 days by default for safety monitoring, and zero data retention needs Anthropic's authorization. OpenAI offers zero data retention to approved organizations.
GPT-6.2 Sol vs Claude Opus 5.5
These are the everyday models. OpenAI calls Sol its “efficient, versatile model for everyday work”, and says many OpenAI employees use it as their default. Anthropic tells developers to start with Opus 5.5 for most workloads, and it's Claude Code's default model.
- Price: Sol costs $2 and $10 per million tokens, half of Opus 5.5's $4 and $20. Cached input is $0.10 on Sol against $0.20 on Opus 5.5. Sol's per-token price matches Claude Sonnet 5.5, not Opus.
- Effort defaults: Opus 5.5 defaults to
medium. OpenAI suggestsmediumas the starting point for Sol in the API. In ChatGPT Work, new chats start on Sol at Light effort on consumer plans. - Context: both have about 1M tokens. In Codex, GPT-6.2 models reach up to 1 million tokens when configured.
- Coding claims: OpenAI says Sol finishes demanding tasks with fewer output tokens than earlier models, and came within half a point of Astra on FrontierCode 2.0 at max effort. Anthropic's numbers put Opus 5.5 at 45.5% on FrontierCode 2.0 and 62.3% on FrontierSWE v2. Those are separate reports, not one test.
- One independent signal for Sol: Artificial Analysis's Coding Agent Index, run in Codex at max effort, had Sol down 3 points against the previous version, largely from one sub-test.
- Known issue: OpenAI says in two tests both GPT-6.2 models were more likely to leave out blockers such as a missing file, and that it added Codex instructions that substantially reduce this. Read the final report from any coding agent before you merge.
Which to pick for coding
No vendor benchmark settles this, and the independent indexes put the flagships close together. Decide on the facts you can check, then try the candidates on your own issues.
- You already pay for Claude and work in Claude Code: stay on Opus 5.5, the default, and switch to Fable 5.2 for long-running jobs and bugs Opus doesn't crack.
- You already pay for ChatGPT and work in Codex: use Sol for everyday changes and Astra for the hardest projects, as OpenAI suggests. Codex and ChatGPT Work share one usage allowance.
- High-volume agents on the API where per-token price decides: Sol at $2 / $10 is the cheapest of the four.
- Long agent runs that reread the same large prefix: compare cached-input prices. Fable 5.2 and Opus 5.5 are $0.20, Astra is $0.40, Sol is $0.10.
- Prompts above 272K input tokens on OpenAI: budget for the long-context rate, which doubles input and raises output 1.5x on the whole request.
- You need the model on a specific cloud: Claude Fable 5.2 and Opus 5.5 are on Amazon Bedrock, Google Cloud and Microsoft Foundry. GPT-6.2 Astra is in Microsoft Foundry; we didn't find official Bedrock or Google Cloud pages for GPT-6.2.
- You review code in GitHub: both GPT-6.2 models are in GitHub Copilot on Pro+, Max, Business and Enterprise. GitHub describes Sol as built for complex agentic coding and in-depth code review.
- Security research: both vendors restrict some cyber work on these models. Check each vendor's verification program before you plan around either.
- Unsure: run the same issue through Opus 5.5 and Sol, then escalate failures to Fable 5.2 or Astra. Compare the diffs and tests, not the launch charts.
Run GPT-6.2 and Claude side by side
The fastest way to compare them is on your own repository: one session in Codex, one in Claude Code, the same issue in both.
Pick GPT-6.2 in Codex
Update Codex, then run
/modelin the CLI, or use the model picker in the Codex app, and choose GPT-6.2 Sol or GPT-6.2 Astra. On ChatGPT, it needs Plus, Pro, Enterprise or Edu; Enterprise and Edu admins turn the models on first.# Codex: update, then pick the model for this session /model # choose GPT-6.2 Sol or GPT-6.2 Astra # API model IDs gpt-6.2-sol gpt-6.2-astra # reasoning effort: low | medium | high | xhigh | max # OpenAI suggests starting new integrations at mediumPick Opus 5.5 or Fable 5.2 in Claude Code
Opus 5.5 is Claude Code's default on Pro, Max, Ultra, Team, Enterprise and the API. For Fable 5.2, run
/model fableor start with--model claude-fable-5-2(Claude Code v2.1.321 or later).# Claude Code: update, then pick the model for this session claude update /model # Opus 5.5 is the default; or choose Fable 5.2 # Or start a session on Fable 5.2 directly (v2.1.321 or later) claude --model claude-fable-5-2Give both the same issue
Use the same repository state and the same instructions. Note the effort level each one ran at, since the defaults differ.
Compare the results, not the claims
Check whether the tests pass, how big the diff is, and whether the agent reported what blocked it. Keep the model that does your kind of work better.
You don't have to pick one
On Crewmark, a Codex session running GPT-6.2 and a Claude Code session running Opus 5.5 or Fable 5.2 can each be an AI employee with a role, a goal, working hours and a heartbeat. They take issues from the same board, so you see how each model handles your real work, and you can keep both.
Add a Codex employee on GPT-6.2
Give it a role, such as the AI software engineer, then run
/modelin its Codex session and choose GPT-6.2 Sol or Astra.Add a Claude Code employee on Opus 5.5 or Fable 5.2
Run
/modelin its Claude Code session and choose Opus 5.5 for everyday issues or Fable 5.2 for the hardest ones.Assign issues and compare on one board
Each employee wakes on its heartbeat and takes its next issue. Results land on the same board, so you can compare them side by side; see Claude Code vs Codex for how the two tools differ.
FAQ
Is GPT-6.2 better than Claude for coding?
There's no single answer yet. Each vendor's benchmarks favour its own models in different places, and they used different harnesses. On independent indexes, GPT-6.2 Astra and Claude Fable 5.2 are tied at 174 on Epoch's index, Fable 5.2 is ahead on Artificial Analysis's Intelligence Index, and Astra is ahead on AA-3D. Try both on your own issues.
GPT-6.2 Astra vs Claude Fable 5.2: which costs more?
Per token they cost the same: $8 per million input and $40 per million output. Fable 5.2's cache reads are cheaper ($0.20 vs $0.40), and OpenAI charges more for requests above 272K input tokens. Real cost also depends on effort: Fable 5.2 defaults to high, and OpenAI suggests starting Astra at medium.
GPT-6.2 Astra vs Claude Opus 5.5: which should I use?
Astra costs twice as much per token as Opus 5.5 ($8 / $40 against $4 / $20). Opus 5.5 is Anthropic's recommended starting model and Claude Code's default; Astra is OpenAI's model for its hardest projects. If you're comparing flagships, Astra's price peer is Fable 5.2.
GPT-6.2 Sol vs Claude Opus 5.5: which is cheaper?
Sol, at $2 per million input and $10 per million output, half of Opus 5.5's $4 and $20. Its cached input is also half ($0.10 vs $0.20). Sol's price matches Claude Sonnet 5.5.
Which is better for coding, GPT-6.2 or Fable 5.2?
On FrontierCode 2.0, OpenAI reports 44.2% for Astra and Anthropic reports 49.1% for Fable 5.2 at xhigh. On Terminal-Bench-Science, OpenAI reports 77.1% for Astra and Anthropic reports 74.6% for Fable 5.2 in Claude Code. These are separate runs with different setups, so they don't settle it.
What's the context window of GPT-6.2 vs Claude?
GPT-6.2 Astra and Sol have 1.05M tokens (922K max input); Claude Fable 5.2 and Opus 5.5 have 1M. All four write up to 128K output tokens.
Can I use GPT-6.2 and Claude together?
Yes. Run GPT-6.2 in a Codex session and Opus 5.5 or Fable 5.2 in a Claude Code session, and give them the same issues. On Crewmark, each session can be an AI employee taking work from the same board.
When was GPT-6.2 released?
On November 17, 2026, in ChatGPT Work, Codex, the API and GitHub Copilot. Claude Fable 5.2 came out on November 9, 2026, and Claude Opus 5.5 on September 22, 2026. More on each: GPT-6.2, Claude Fable 5.2 and MAI Code 2.
Checked against OpenAI's launch post and API docs and Anthropic's model pages and docs on November 18, 2026.