Claude Opus 5.5 vs GPT-6.1 Sol for Agentic Coding

Claude Opus 5.5 is the stronger pick on the agentic coding benchmarks its maker published, and Anthropic says it costs 40% less to run than Opus 5. GPT-6.1 Sol is the cheaper model, at half Opus 5.5's list price per token and half its cached-input price, and OpenAI says it nearly matches its larger GPT-6 Astra on agentic coding. Neither vendor has published the other's newest model on the same coding test, so the honest answer is to run both on your own tasks. This page shows what each vendor claims, what the claims don't tell you, and how to run a fair test in an afternoon.

Anthropic released Opus 5.5 on September 22, 2026. OpenAI announced GPT-6.1 Sol at DevDay a week later, on September 29. We checked every figure below against first-party sources on October 9, 2026.

Specs side by side#

	Claude Opus 5.5	GPT-6.1 Sol
Released	September 22, 2026	September 29, 2026
Vendor's pitch	Built for long-running agentic coding and knowledge work	Nearly matches GPT-6 Astra on agentic coding at one-fifth of Astra's price
Price per 1M tokens (input / output)	$4 / $20	$2 / $10
Cached input per 1M tokens	$0.20	$0.10
Context window	1M tokens	1,050,000 tokens (922,000 input)
Max output	128K tokens	128K tokens
Reasoning	Adaptive thinking, always on; effort defaults to medium	Configurable reasoning effort
Knowledge cutoff	June 2026	April 30, 2026
API model ID	claude-opus-5-5	gpt-6.1-sol
Where to use it	Claude apps and Claude Code; Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry	Codex and ChatGPT Work (Plus and above); OpenAI API

Two pricing details matter for agents. GPT-6.1 Sol charges double for input and 1.5 times for output on the whole request once a prompt passes 272K input tokens (OpenAI), which long-context agents can hit. And Opus 5.5 can't run with thinking turned off: a request that disables thinking returns an error (Anthropic), which matters if you call it directly from your own harness.

What each vendor reports#

Anthropic published Opus 5.5 against its own models, GPT-6 Astra, and GPT-5.6 Sol (announcement). On Terminal-Bench 4.0, Opus 5.5 scored 66.4%, ahead of Claude Fable 5.1 at 55.8%, Opus 5 at 52.3%, and GPT-6 Astra at 57.9% as reported by OpenAI. On FrontierCode v1.1 the gap was narrower: 54.4% for Opus 5.5 against 53.3% for Astra. Opus 5.5 ran at its xhigh effort setting for these results, and Anthropic says that at default settings it costs 40% less than Opus 5 on typical workloads.

OpenAI published GPT-6.1 Sol against Astra, the previous GPT-6 Sol, and Opus 5.5 (announcement). Its headline claim is near-Astra performance on agentic coding, computer use, and professional work at one-fifth of Astra's standard prices. On Terminal-Bench Science 0.1 at maximum effort, OpenAI reports GPT-6.1 Sol costing $5.47 per task on average, against $23.21 for Opus 5.5 and $23.80 for Astra, with Astra still scoring highest.

What that leaves out. Anthropic's table was published before GPT-6.1 Sol existed, and OpenAI's coding comparisons are mostly against its own models. Benchmarks also run inside a particular harness and effort setting, and your results will depend on your repository, your instructions, and your checks as much as on the model. Treat both tables as reasons to test, not as answers.

Cost per task, not cost per token#

List prices make Sol look half as expensive, and for many workloads it will be. Three things move the real number:

• Tokens per task. A model that spends fewer tokens or needs fewer retries can cost less per finished change despite a higher price. Anthropic says Opus 5.5 uses fewer tokens per task than Opus 5; measure both models on your own work.
• Cache reads. Agents re-read the same instructions and files on every turn, so cached input is often the bulk of the bill. Sol's $0.10 cached rate is half of Opus 5.5's $0.20.
• Effort settings. Both models spend far more at their highest effort levels. Run your test at the effort you'd actually use day to day.

If you run Claude Code or Codex on a subscription rather than API keys, the question becomes which model gets more done before you hit your plan's limits. Agentastic's Usage & Cost view shows quota and API-equivalent cost for both from logs already on your disk.

How to test both on your own repository#

A fair test takes an afternoon:

1. Pick three to five real tasks from your backlog: a bug with a reproduction, a small feature, a refactor, a flaky test. Skip toy problems.
2. Write each task once in a file, with the same finish line for both models, such as "done when npm test and npm run lint pass."
3. Start both from the same commit, in separate worktrees, so neither sees the other's changes. In Agentastic:

Without Agentastic, create two worktrees with git worktree add and run claude --model claude-opus-5-5 in one and codex -m gpt-6.1-sol in the other.

4. Score the results: do tests pass, how many lines did you change before merging, how many review comments did each get, how long did it run, and what did it cost.
5. Repeat on the next task. One task proves nothing; five start to show a pattern.

Running two agents from different vendors side by side is what Agentastic is built for: each gets its own worktree and terminal or chat, and the diff viewer and review passes make the comparison quick. See Claude Code vs Codex for the differences between the agents themselves.

Which to choose#

• Choose Opus 5.5 for long, ambiguous, multi-step work where getting it right the first time saves more than the token price, and if your team already works in Claude Code.
• Choose GPT-6.1 Sol for high-volume work: triage, CI fixes, test generation, and loops that re-read the same context many times a day, and if your team already works in Codex.
• Use both when the work allows it: route by task type, or have one model review the other's diff. A second model is one of the cheapest checks you can add.

For the hardest problems each vendor points to its larger model: OpenAI to GPT-6 Astra, and Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1 on most work.

Sources#

• Anthropic: Introducing Claude Opus 5.5, Opus 5.5 overview, what's new in Opus 5.5
• OpenAI: Introducing GPT-6.1 Sol, GPT-6.1 Sol model page, DevDay 2026 recap
• Simon Willison, Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war
• On this site: Claude Code vs Codex, how to compare coding agents, track Claude Code and Codex usage

Frequently asked questions#

Is Claude Opus 5.5 better than GPT-6.1 Sol for coding?#

On the agentic coding benchmarks Anthropic published, Opus 5.5 leads, including 66.4% on Terminal-Bench 4.0. OpenAI hasn't published GPT-6.1 Sol on the same tests, and positions Sol as nearly matching its larger GPT-6 Astra at a fraction of the price. There is no neutral head-to-head yet, so test both on tasks from your own backlog.

Which is cheaper, Opus 5.5 or GPT-6.1 Sol?#

GPT-6.1 Sol has the lower list price: $2 per million input tokens and $10 per million output tokens, against $4 and $20 for Opus 5.5. Cached input is $0.10 per million for Sol and $0.20 for Opus 5.5. Cost per finished task also depends on how many tokens each model spends, so measure it on your own work.

Can I use GPT-6.1 Sol in Codex?#

Yes. OpenAI made GPT-6.1 Sol available in Codex and ChatGPT Work for Plus, Pro, Business, Enterprise, and Edu users on September 29, 2026, and in the API as gpt-6.1-sol. It wasn't yet available in ChatGPT's Chat mode at launch.

Can I use Claude Opus 5.5 in Claude Code?#

Yes. Opus 5.5 is available in Anthropic's apps and on the Claude API as claude-opus-5-5, so you can pick it in Claude Code with /model or start a session with claude --model claude-opus-5-5. It is also on Amazon Bedrock, Google Cloud, and Microsoft Foundry.

What is the context window of Opus 5.5 and GPT-6.1 Sol?#

Both accept about a million tokens. Opus 5.5 has a 1M-token context window and up to 128K output tokens. GPT-6.1 Sol has a 1,050,000-token context window with up to 922,000 input and 128,000 output tokens, and prompts above 272K input tokens cost more per token.

How do I compare Opus 5.5 and GPT-6.1 Sol on my own code?#

Pick three to five real tasks, give both models the same prompt and repository instructions, and run them in separate git worktrees from the same commit so neither sees the other's changes. Compare whether tests pass, how much you had to fix, how long each took, and what each cost.