Claude Opus 5.5 vs GPT-6 Sol: One Shared Benchmark, Two Price Sheets

Claude Opus 5.5 and GPT-6 Sol launched on the same day, Tuesday 22 September 2026, and neither vendor’s launch post measures its model against the other’s. The one test both report is FrontierCode v1.1 Main, where the benchmark’s own leaderboard has Opus 5.5 at 54.6% and GPT-6 Sol at 45.9% at their default medium effort, for about the same spend. On list price, GPT-6 Sol costs exactly half as much per token, up to 272,000 input tokens.

This page was researched and drafted with the help of Claude, an Anthropic model family that includes one of the two models compared, so every figure here is sourced to a vendor page or an independent evaluator, and neither vendor’s self-reported numbers are presented as independent. We name no overall winner.

Is there a head-to-head benchmark for Opus 5.5 and GPT-6 Sol?

Only one so far: FrontierCode v1.1 Main, a coding benchmark maintained by Cognition, which added both models to its public leaderboard on 22 September 2026. Read on 23 September, the board has Opus 5.5 at 54.6% and GPT-6 Sol at 45.9% at their default medium effort, with a mean spend per rollout of $0.80 and $0.77.

The rows ran in different harnesses, which Cognition names as claude-code for Opus 5.5 and codex for GPT-6 Sol, so this partly tests each vendor’s own coding agent. On best runs, the board ranks Opus 5.5 first and GPT-6 Sol seventh.

Opus 5.5 gains nothing above medium: max costs nearly eight times as much per rollout and scores 0.2 points lower. GPT-6 Sol improves at every step, and its max run costs about a third of Opus 5.5’s. It is one benchmark, measuring one kind of work.

What do Anthropic and OpenAI each claim?

Each vendor reports its own set of tests, and apart from FrontierCode they do not overlap. Anthropic’s headline figures for Opus 5.5 include 66.4% on Terminal-Bench 4.0; OpenAI’s for GPT-6 Sol include 68.8% on DeepSWE v1.1. Anthropic’s table leaves GPT-6 Sol out, OpenAI’s charts compare against Claude Opus 5, Fable 5.1 and Fable 5, and every figure here is vendor-reported.

Anthropic’s figures are at max effort unless marked; CursorBench falls to 52.5% at the default medium. Anthropic also ran its benchmarks with production safeguards on, and says that when they intervened, Opus 4.8 finished cyber tasks and Opus 5 finished bio and frontier-LLM-development tasks. Some Opus 5.5 scores therefore include work done by other models.

Both posts also report OSWorld 2.0 scores, left out here on purpose: the same older model, Opus 5, scores 74.0% in Anthropic’s table and 70.2% in OpenAI’s chart, so the setups differ.

What do independent evaluators say?

Artificial Analysis, which runs both models through its own harness, scores Opus 5.5 at 58 and GPT-6 Sol at 48 on its Intelligence Index v4.3.2, both at max effort, and puts the cost per Intelligence Index task at $5.98 against $1.06. Opus 5.5 leads on all ten component evals shown on its comparison page, three of them by a single point. Zapier has also scored both models; ARC Prize, so far, only Opus 5.5.

The same data shows Opus 5.5 used 260 million output tokens across the index, GPT-6 Sol 77 million. Artificial Analysis puts GPT-6 Sol level with GPT-5.6 on the index at about half the cost per task, with its GDPval-AA score down about 100 Elo. The index was revised four times in September 2026, and in our earlier comparison of GPT-6 Astra and Claude Fable 5.1 we watched both headline scores move within days. Treat the version number as part of the score.

Zapier, which runs AutomationBench, has both models on its 1.0.6 leaderboard, priced the same way. Opus 5.5 reaches 40.0% at max for $1.28 a task, 34.4% at xhigh for $0.80 and 32.0% at high for $0.65. GPT-6 Sol reaches 33.2% at xhigh for $0.27 and 32.0% at max for $0.34. GPT-6 Astra tops the board at 41.4%. Anthropic notes that Zapier ran without fallback models, so any safeguard intervention counted as a failure.

ARC Prize has a verified page for Opus 5.5, dated 22 September: 98.5% on ARC-AGI-1 and 93.3% on ARC-AGI-2’s semi-private set, both at high effort, with no ARC-AGI-3 score. It has no GPT-6 Sol page yet. On 23 September, neither model was ranked on arena.ai’s text leaderboard or listed on the official Terminal-Bench 4.0 leaderboard at tbench.ai.

How much do Opus 5.5 and GPT-6 Sol cost per token?

GPT-6 Sol lists at exactly half Opus 5.5’s per-token price: $2 per million input tokens and $10 per million output, against $4 and $20. Cache reads cost $0.20 per million on both. That holds only up to 272,000 input tokens; above it, OpenAI bills the whole Sol request at $4 input, $0.40 cached and $15 output, while Anthropic charges Opus 5.5 one rate across its full 1M-token window.

Opus 5.5’s fast mode is a research preview, on Anthropic’s own API only and not with Batch. Its cache reads are 0.05 times the input price; Anthropic’s other models charge 0.1 times. Reasoning tokens bill as output on both, as our July breakdown of true LLM cost explains.

Both discount claims use each vendor’s own earlier price. OpenAI’s 50% cut is against GPT-5.6 Sol’s promotional $4 and $20, which OpenAI says runs at least through 21 November 2026. Anthropic says Opus 5.5 at default settings will cost 40% less than Opus 5 ($5 and $25) on typical workloads; we have not tested that.

What does a real workload cost on each model?

With identical token counts, Opus 5.5 costs twice as much as GPT-6 Sol for ordinary prompts and 1.75 times as much for a cache-heavy agent turn, or $7.00 against $4.00 over 50 turns. Above 272,000 input tokens the gap almost closes, and a repeat question against a large cached document is cheaper on Opus 5.5.

One extra input token past 272,000 nearly doubles the Sol bill, from $0.594 to $1.163.

The equal-token assumption is the weak point. Neither vendor publishes how the same text tokenizes on the other’s model, and the two spend very different amounts of output. Measured per task instead, the gap runs from roughly level to more than fivefold:

• FrontierCode at medium: $0.80 against $0.77 per rollout.
• AutomationBench at near-equal scores: Opus 5.5 at high (32.0%) for $0.65, GPT-6 Sol at xhigh (33.2%) for $0.27, about 2.4 times.
• FrontierCode at max: $6.19 against $2.07, about three times.
• Artificial Analysis index at max: $5.98 against $1.06, about 5.6 times.

What API differences will developers notice?

The one most likely to bite is that Opus 5.5 cannot run with thinking switched off. Anthropic lists its thinking as adaptive and always on, with medium as the default effort, while GPT-6 Sol accepts a reasoning effort of none. Sol also caps input at 922,000 tokens inside its 1,050,000-token window.

• On Chat Completions, GPT-6 Sol supports function calling only at reasoning effort none.
• Anthropic says most cybersecurity tasks sent to Opus 5.5 will be re-routed to Opus 4.8. OpenAI’s Preparedness assessment treats GPT-6 Sol as High capability in cybersecurity and in biological and chemical, in an appendix added to the GPT-6 Astra system card on 22 September.
• Preserved thinking, which ties Opus 5.5’s reasoning blocks to the model and conversation that produced them, is enforced for API accounts created on or after 31 August 2026. Zero data retention is available.
• Opus 5.5 is on the Claude Platform, AWS, Google Cloud and Microsoft Azure; GPT-6 Sol is in the API, in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users (not yet in Chat), and in GitHub Copilot since 22 September.

What should you do?

Test on your own workload; no published number settles it.

• Set effort explicitly on both. On FrontierCode, Opus 5.5 scored as well at medium as at max for about an eighth of the spend, while GPT-6 Sol gained at every step, so test Sol at more than one setting.
• Count billed tokens, not prompt length: each model’s tokenizer and reasoning set the real bill.
• Check where your prompts sit against 272,000 tokens. Above it, GPT-6 Sol’s price advantage mostly disappears, and repeated queries over one large cached document cost less on Opus 5.5 at list price.
• Pick Sol if you need calls with no reasoning at all. Opus 5.5 does not offer that mode.
• Re-check around 7 October 2026, when the independent boards may have moved or caught up.