GPT-6 Sol vs Claude Opus 5.5 for Coding (2026)

Claude Opus 5.5 is the better coding model. GPT-6 Sol is the cheaper one, and on routine work it is cheap enough that the gap stops mattering. Both shipped on September 22, 2026, and neither vendor benchmarked against the other, so the only clean comparison comes from independent testing. Route hard, terminal-heavy work to Opus 5.5 in Claude Code and high-volume, well-scoped work to Sol in Codex.

What each model actually is

GPT-6 Sol is OpenAI's mid-tier GPT-6 model, sitting under GPT-6 Astra (the flagship, released earlier in September) and above GPT-6 Luna (the budget tier). OpenAI's launch post says Sol was trained with methods similar to Astra's and positions it for "complex coding and agentic workflows." The headline is price: OpenAI cut Sol to $2 input and $10 output per million tokens, half of GPT-5.6 Sol's promotional $4/$20. It ships in Codex and ChatGPT Work, and in the API as gpt-6-sol.

Claude Opus 5.5 is the first model in Anthropic's Claude 5.5 family, with Sonnet 5.5 and Haiku 5.5 promised "in the coming weeks." Per Anthropic's announcement, it scores above Claude Fable 5.1 on every benchmark in Anthropic's own table, costs $4/$20 (20% below Opus 5), cuts cache reads 60% to $0.20, and generates output more than 30% faster than Opus 5. It is the default model in Claude Code and is available as claude-opus-5-5 in the API, on Amazon Bedrock, Google Cloud, and Microsoft Foundry.

So this is not a flagship vs flagship fight. Sol is OpenAI's cost-efficiency model. Opus 5.5 is Anthropic's cost-reduced flagship. That framing explains almost every number below.

The benchmark problem: nobody compared these two

Here is the thing most launch-day coverage glossed over. OpenAI's charts compare Sol against Claude Opus 5 and Fable 5.1. Anthropic's table compares Opus 5.5 against GPT-6 Astra and GPT-5.6 Sol. Neither lab put the other's same-day model on its chart, as Digital Applied pointed out.

That leaves two kinds of evidence: the few benchmarks both vendors happen to report, and independent runs on the same harness.

Vendor-reported scores on shared benchmarks

Sources: Anthropic for Opus 5.5, OpenAI and Kingy AI's summary of OpenAI's charts for Sol.

FrontierCode looks like a 5-point Opus win, but treat it as inconclusive. The two vendors disagree about the same older model on that benchmark: Anthropic lists Opus 5 at 48.0%, OpenAI's chart lists it at 53.4%. A five-point gap between two runs of the same model is as large as the gap between Sol and Opus 5.5.

The one Anthropic claim that touches the Sol line is CursorBench 4.0, where Opus 5.5 scored 57.8% against GPT-5.6 Sol's 41.7%, "for about a third of the cost per task." That is the previous Sol, not GPT-6 Sol, so it tells you where OpenAI started from, not where it landed.

Independent scores on the same harness

Artificial Analysis ran both models across its ten-test Intelligence Index at every effort level, with costs measured at list prices including cache reads and writes. This is the cleanest head-to-head available.

Figures as read by Digital Applied from Artificial Analysis on September 22, 2026.

Three things stand out.

1. Terminal work is the widest gap. Opus 5.5 at its default beats Sol at max by 8.6 points on Terminal-Bench 4.0, and Opus 5.5 climbs to 59.6% at xhigh. Sol drops off fast as effort falls: 30.3% at xhigh, 26.3% at high. If your agent lives in a shell, running tests and chaining commands, this is the number that matters.
2. Business automation is a tie. On AutomationBench-AA, Sol at xhigh matches Opus 5.5 at medium for about 40% of the cost per task.
3. Sol at max is barely cheaper. It costs $1.06 per index task against Opus 5.5's $1.34, and it uses more output tokens (31.2k vs 25.7k) to get there. At the top of the effort ladder, the per-token discount mostly gets eaten by extra reasoning.

Pricing: the full rate card

Sources: OpenAI's GPT-6 Sol model page, Anthropic, and Digital Applied's Opus 5.5 breakdown.

Two lines break the "Sol is half price" story, and both matter more for coding agents than for chat.

Cache reads are identical. A coding agent re-sends the same repo context, system prompt, and conversation history on every step. Most of its input tokens are cache hits. At $0.20 per million on both sides, the biggest line on an agent bill does not get cheaper when you switch to Sol.

The 272K cliff. Opus 5.5 charges standard rates across the full 1M window. Sol bills the entire request at 2x input and cache rates plus 1.5x output once input passes 272K tokens. Agents working in large monorepos, or long sessions that never compact, cross that line more often than you would guess.

Cost per task with cache reads: a worked example

Token counts below are illustrative and held equal for both models, so the only variable is the rate card. The shape is a typical mid-size agent task: lots of cached context, modest fresh input, some cache writes, and reasoning-heavy output.

At equal tokens, Sol costs 56% of Opus 5.5, not 50%. Push every request past 272K and Sol costs exactly the same as Opus 5.5 on this task.

Real token counts are not equal, which is why measured cost per task beats list price. At default (medium) effort, Artificial Analysis measured Sol at $0.25 per index task against $1.34 for Opus 5.5, roughly a fifth. But that buys an index score of 39.8 against 51.2. Every Sol setting below max costs less than the cheapest Opus 5.5 setting, and everything Opus 5.5 does from medium upward scores higher than anything Sol reaches. The real question is which price band a task belongs in, not which model is "better."

One counterintuitive result from the same data: Sol with reasoning set to none scored 28.1 for $0.33 per task, lower and more expensive than Sol at low (33.9 for $0.13). A model that does not think takes more steps, and every step re-sends the context. Turning reasoning off is not a reliable way to save money on agent work.

Harness: Codex vs Claude Code

Most people do not pick a model. They pick a harness and use its default. That makes the harness half of this comparison.

GPT-6 Sol runs in Codex. OpenAI rolled Sol and Luna into Codex and ChatGPT Work on launch day for Plus, Pro, Business, Enterprise, and Edu users. Codex does not upgrade the model on its own, so existing setups need the model switched to gpt-6-sol (the /model command works mid-session). Reasoning effort is set with model_reasoning_effort in your Codex config.

Claude Opus 5.5 runs in Claude Code. It became the Claude Code default on launch day, and Anthropic raised five-hour session limits by 20% for Pro, Max, and Team subscribers from September 22. Because Opus 5.5 is also cheaper per token, subscription usage stretches further inside those limits. The details, including the new medium effort default, are in our Opus 5.5 in Claude Code guide.

The harness-level differences that actually change your workflow:

The cache behavior is the sleeper issue. If you build a router that raises effort on hard turns and drops it on easy ones, Sol keeps its cache warm across those switches. On Opus 5.5, the same pattern rebuilds the cache on every change unless you opt into the per-message effort beta. If you run a custom agent on the API, test this before trusting your cost projections.

Opus 5.5 also has a silent breaking change worth knowing if you run your own harness: short notes between tool calls now come back as empty thinking blocks instead of visible text. Anything that parsed those notes will quietly stop working.

For the broader tool comparison (hooks, subagents, remote runs, rate-limit history), see Claude Code vs Codex. The harness gap there has not closed just because the models changed.

Routing rules

If you can use both, route by price band and task shape. These rules follow directly from the independent data above.

1. Terminal-heavy coding goes to Opus 5.5 at medium. Running tests, debugging a failing build, multi-step shell work. It leads Sol at max by 8.6 points on Terminal-Bench 4.0 at about a quarter more per index task. Step up to xhigh only on tasks that failed at medium.
2. Budget, high-volume coding goes to Sol at medium to xhigh. Test scaffolding, lint fixes, small refactors, codemod-style edits across many files. Below roughly $0.55 per task, Sol scores higher than any Opus 5.5 setting.
3. Cross-app automation starts on Sol at xhigh. Ticket triage, CRM updates, glue workflows. It ties Opus 5.5 at medium on AutomationBench-AA for about 40% of the cost.
4. Anything over 272K input per request goes to Opus 5.5. The long-context surcharge erases Sol's price advantage, and Opus 5.5 does not charge more for the full window.
5. Do not run Sol at max as a default. It costs about as much as Opus 5.5 at medium and still scores lower on coding.
6. Do not turn Sol's reasoning to none for agent loops. It costs more per task than low and scores worse.
7. Validate on your own repo. Take 10 to 20 real tasks, run Sol at xhigh and Opus 5.5 at medium, and compare cost per completed task plus human correction time. One independent harness is a strong signal, not a verdict on your codebase.

Who each one is for

Pick Claude Opus 5.5 if your work is agentic coding in a terminal, you ship production changes where a missed edge case costs more than the tokens, you already live in Claude Code, or your sessions routinely carry huge context. It is the stronger coding model at every effort level from medium up, and it does not punish long prompts.

Pick GPT-6 Sol if you run high-volume, well-defined tasks where cost per task decides whether the workflow exists at all, you are already standardized on Codex or ChatGPT Work, or you need fine-grained control over reasoning, including cache-safe effort switching. It is the best cheap model in this matchup by a wide margin.

When to pick the other one

Being fair to both sides:

Choose Sol even if you love Claude Code when the task is repetitive and the bar is "correct and cheap," not "best possible." At its default, Sol costs about a fifth of Opus 5.5 per index task. For bulk migrations, generated tests, and automation that runs thousands of times a day, that difference funds a lot of retries. Sol also knows less but guesses less: on AA-Omniscience, Opus 5.5 at medium answers more questions correctly (64.5% vs 54.5%) but, when it does not know, gives a wrong answer 68.4% of the time against 60.1% for Sol at max. Neither is low enough to skip verification, but Sol is less confidently wrong.

Choose Opus 5.5 even if you love Codex when the task is hard, long, or lives in a shell. Sol's top score on the Intelligence Index (47.5 at max) sits below Opus 5.5 at its default setting (51.2). Paying more per token for a model that finishes in fewer steps and fewer human corrections is often the cheaper option per shipped change.

Anthropic's own launch post said it plainly: at this level of capability, benchmark margins have become a less reliable guide to real-world differences. The gap between these two is real on terminal coding and close to zero on business automation. Your repo is the benchmark that counts.

Posted by @speedy_devv