Claude Opus 5.5 vs GPT-6 Astra
Claude Opus 5.5 beats GPT-6 Astra on every benchmark both have published and costs less than half as much per token: $4 input and $20 output per 1M, against $10 and $50. Our feature session comes to $21.60 on Opus 5.5 and $54 on Astra.
The scores
Terminal-Bench 4.0, agentic coding in a real terminal: Opus 5.5 66.4% at xhigh effort, Astra 57.9%, a figure OpenAI reported at high effort and Anthropic quoted. FrontierCode 1.1: 54.4% against 53.3%, close enough to call a tie. Humanity's Last Exam with tools: 67.7% against 57.2%.
What a session costs
Our feature session costs $21.60 on Opus 5.5 and $54 on Astra. With 90% of prompts read from the cache it is $5.18 against $15.12, a wider gap, because cached input on Astra costs $1 against $0.20 on Opus. A long agent day is $195 against $488.
Divided by the Terminal-Bench score, the feature session costs $0.33 per point on Opus 5.5 and $0.93 on Astra.
Dates and size
GPT-6 Astra was first listed on the API on 4 September 2026; Anthropic released Opus 5.5 three weeks later, on 22 September. Astra reads up to 1.05M tokens and Opus 5.5 up to 1M, and both write up to 128K. Claude models run in Claude Code and OpenAI's in Codex; Claude Code vs Codex compares the tools themselves.
Side by side
| Claude Opus 5.5 | GPT-6 Astra | |
|---|---|---|
| Maker | Anthropic | OpenAI |
| Released | 22 September 2026[1] | 4 September 2026[2] first API listing |
| Context window | 1M tokens[1] | 1.05M tokens[3] |
| Max output | 128K tokens[1] | 128K tokens[3] |
| Input price | $4 per 1M tokens[1] | $10 per 1M tokens[3] |
| Output price | $20 per 1M tokens[1] | $50 per 1M tokens[3] |
| Cached input | $0.20 per 1M tokens[4] | $1 per 1M tokens[3] |
| Long prompts | One price | One price |
| Price change | None announced | None announced |
| Open weights | No[1] Offered through the Claude API and cloud platforms; no weights are published. | No[5] Listed as an API model, not in OpenAI's open-weight section. |
| Reads | text, image[1] | text, image[3] |
Benchmarks
Scores as the makers published them, each with its source and date. Makers run these at different settings, so read a gap of a few points as a tie.
| Benchmark | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|
| Terminal-Bench 4.0Agentic coding tasks in a real terminal. | 66.4%[6] Run at xhigh effort. | 57.9%[6] As reported by OpenAI (high effort), quoted by Anthropic. |
| FrontierCode 1.1 (Main)Hard software engineering tasks. | 54.4%[6] Max effort. | 53.3%[6] Quoted by Anthropic. |
| Humanity's Last Exam (with tools)Expert-level questions across many fields, tools allowed. | 67.7%[6] | 57.2%[6] Quoted by Anthropic. |
What a coding session costs
Our own estimate from the list prices above, for three sessions we define: a coding agent resends the conversation with every call, so prompts are large and grow. "Cached" assumes 90% of prompt tokens are read from the prompt cache and leaves out cache write fees. A long-prompt price applies when a call's prompt is over its threshold. Work out your own mix with the token cost calculator.
| Session | Claude Opus 5.5 | Claude Opus 5.5 cached | GPT-6 Astra | GPT-6 Astra cached |
|---|---|---|---|---|
| Quick fix10 calls of 25K prompt and 1.5K output tokens | $1.30 | $0.45 | $3.25 | $1.23 |
| Feature session60 calls of 80K prompt and 2K output tokens | $21.60 | $5.18 | $54 | $15.12 |
| Long agent day300 calls of 150K prompt and 2.5K output tokens | $195 | $41.10 | $488 | $123 |
Value for money and release gap
Claude Opus 5.5 is dated 3 weeks later (18 days; GPT-6 Astra: first API listing). Price per point divides our feature-session cost by a published score: lower means more score for the money. It is a rough lens, not a ranking.
| Measure | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|
| Input tokens per $1 | 250K | 100K |
| Filling the whole context window once | $4.00 | $10.50 |
| Feature session per point of Terminal-Bench 4.0 | $0.33 | $0.93 |
| Feature session per point of FrontierCode 1.1 (Main) | $0.40 | $1.01 |
| Feature session per point of Humanity's Last Exam (with tools) | $0.32 | $0.94 |
Questions
Is Claude Opus 5.5 better than GPT-6 Astra?
On published scores, yes on all three: 66.4% against 57.9% on Terminal-Bench 4.0, 54.4% against 53.3% on FrontierCode 1.1 and 67.7% against 57.2% on Humanity's Last Exam with tools.
How much more does GPT-6 Astra cost than Opus 5.5?
2.5 times as much per token: $10 against $4 per 1M input tokens and $50 against $20 per 1M output. Our feature session costs $54 on Astra and $21.60 on Opus 5.5.
Are the Opus 5.5 and GPT-6 Astra scores comparable?
Only roughly. The Astra Terminal-Bench figure is OpenAI's own at high effort, quoted by Anthropic, and Opus 5.5 ran at xhigh effort. A gap of a point or so, as on FrontierCode, is a tie.
Most compared
Any other two models: pick them in the lists at the top of the page.
- Claude Opus 5.5 vs Claude Sonnet 5.5
- Claude Fable 5.1 vs Claude Opus 5.5
- Claude Opus 5 vs Claude Opus 5.5
- Claude Opus 5.5 vs GPT-6.1 Sol
- Claude Opus 5.5 vs GPT-5.6 Sol
- Claude Fable 5.1 vs GPT-6 Astra
- GPT-6.1 Sol vs GPT-6 Astra
- Claude Opus 5.5 vs Gemini 4 Argon
- Claude Opus 5.5 vs Gemini 3.1 Pro Preview
- Claude Opus 5.5 vs Grok 4.7
- Claude Opus 5.5 vs DeepSeek V4 Pro
- Claude Sonnet 5 vs Claude Sonnet 5.5
- Claude Haiku 5.5 vs Claude Sonnet 5.5
- Claude Haiku 4.5 vs Claude Haiku 5.5
- Claude Fable 5 vs Claude Fable 5.1
- Claude Sonnet 5.5 vs GPT-6.1 Sol
- Claude Haiku 5.5 vs GPT-6 Luna
- GPT-6.1 Sol vs GPT-6 Sol
- Claude Sonnet 5.5 vs Gemini 4 Argon
- Gemini 4 Argon vs GPT-6.1 Sol
- Gemini 3.1 Pro Preview vs Gemini 4 Argon
- Claude Haiku 5.5 vs Gemini 3.8 Flash
- Claude Sonnet 5.5 vs Kimi K3
- Claude Sonnet 5.5 vs GLM-5.3
More tools
Sources
Numbers from the makers' own pages, checked by hand, last read . List prices in US dollars before tax. Spotted a mistake? office@shuriken.sh.
Made by Shuriken, the harness for coding agents. Its hot swapping moves your chats to another of your accounts or API keys when one nears its limit. Use only accounts you own and are entitled to use.
Try Shuriken on your own computerYou're signed in
Add your Claude, Codex and Gemini accounts once. Shuriken switches between them so your agents keep working.Your accounts, projects and chats are in the app.
Coming soon