Claude Opus 5.5 vs GPT-6.1 Sol: Level at Default Effort, No Contest Above It

OpenAI shipped GPT-6.1 Sol at DevDay on 29 September 2026, seven days after GPT-6 Sol and eight days after Anthropic's Claude Opus 5.5. The pitch is "near-Astra intelligence for a fifth of the price," but the more useful comparison is against the model it is actually priced against — because at $2 / $10 it is exactly half of Opus 5.5's $4 / $20, and Artificial Analysis has run both models across all ten effort levels in a single harness.

That independent run is what this article is built on, because the two vendors benchmarked against different things. OpenAI's charts compare Sol against Opus 5.5 on four evaluations it chose; Anthropic's launch table does not include GPT-6.1 Sol at all. The one harness that ran both on the same ten evaluations at every effort setting points somewhere neither lab emphasised.

Two releases, eight days apart

The timing matters because three models landed inside a week. Claude Opus 5.5 shipped on 22 September, Claude Sonnet 5.5 on 28 September, and GPT-6.1 Sol on 29 September at OpenAI's DevDay. Sol's list price — $2 in, $10 out — is identical to Sonnet 5.5's, which makes the mid-tier the contested ground and Opus 5.5 the model both are undercutting.

GPT-6.1 Sol is not a new architecture. It is a point release that keeps GPT-6 Sol's standard rates, halves the cached-input rate to $0.10, and buys four points of Intelligence Index. Artificial Analysis's summary of the replacement is pointed: "GPT-6.1 Sol replaces GPT-6 Sol after just 7 days."

One correction worth making early. Several third-party trackers still list Opus 5.5 at $5 / $25 — that is Opus 5's card. Anthropic's own pricing page and model documentation put Opus 5.5 at $4 input, $0.20 cached input, $20 output, with cache writes at $5 for five minutes or $8 for an hour. Every calculation below uses the published Opus 5.5 rates.

The ladder: all ten settings, one harness

This is the table the comparison should be built on. Both models ran the identical ten evaluations under Artificial Analysis — AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience and AA-LCR v1.1 — and cost is measured per task including cache reads and writes.

Where the ladders cross, and where they stop

Read the ten rows in cost order and three findings fall out of the table without any interpretation.

1. At the floors, the scores are identical

Both models score 42 at low effort. Opus charges $0.55 a task for it; Sol charges $0.13. That is the cleanest single cost comparison in the release, because there is no capability argument to have — the scores match exactly and the price differs by 76%.

2. At the defaults, Sol is ahead on both axes

Opus 5.5's default is medium: 51 for $1.34. GPT-6.1 Sol's max setting is 52 for $0.72 — a point higher for 46% less. Sol's own default, medium, is 48 for $0.21, which is a different conversation: 84% cheaper than Opus with a three-point capability gap. The practical read is that Sol's top setting beats Opus's middle setting on both score and price.

3. Above index 51, Sol has nothing

This is the finding the vendor charts miss. Opus 5.5 at high, xhigh and max scores 54, 56 and 58. Sol's ceiling is 52, reached only at max, and it costs $0.72 to get there. Paying Sol more does not buy more: the four-step climb from xhigh to max costs 85% more per task for a single index point.

So the routing question is not "which is better." It is "does the task need an index score above 51?" If no, Sol is the cheaper answer at every level. If yes, Opus 5.5 is the only answer, and the premium starts at $1.82 a task.

A note on the numbers moving. Artificial Analysis has revised these figures since the GPT-6 Sol launch — Opus 5.5's ceiling and default both now read one point higher than in the earlier snapshot, and output speeds have been re-measured. The index is a live product; re-check the model pages before making a routing commitment on a single point of difference.

Test by test: sixteen components

The composite hides a lot. At maximum effort on both models, here is every component score Artificial Analysis publishes, with the leader marked.

Opus 5.5 wins thirteen of sixteen and ties one. Its margins are widest exactly where you would expect a $4/$20 frontier-class model to win: knowledge-work documents (+291 Elo on GDPval-AA, +243 on AA-Briefcase), scientific coding (+12.7 on SciCode) and expert reasoning (+8.5 on Humanity's Last Exam).

Sol's one clear win is GDP.pdf, the Surge AI benchmark over dense professional PDFs — financial statements, legal filings, multi-column tables. Sol scores 31.0% at max and 32.0% at high, against Opus 5.5's 26.2% at max. OpenAI's own chart puts its figure at 32.0% against Opus 5.5 at 28.8% "with fallbacks," and against GPT-6 Astra's 32.2%. For document-extraction pipelines, that is a real and specific advantage — and it is the only one.

An oddity in both models: more effort is not always better

GDP.pdf peaks in the middle of the ladder for both models. Opus 5.5 is strongest at high (28.8%) and weaker at max (26.2%). Sol is strongest at high (32.0%) and weaker at max (31.0%). And on AutomationBench-AA, Sol's best result is at xhigh (66.6%), above its own max (64.9%).

OpenAI's own coding chart shows the same shape more dramatically, which is worth stating precisely because it determines the setting you should ship:

For Sol, the operational instruction is unambiguous: ship high for coding, medium for business workflows, and treat max as unproven on your workload. The same warning applies with less force to Opus 5.5, where GDP.pdf regresses at max even as the composite index keeps climbing.

The 272K cliff

GPT-6.1 Sol's price advantage is usually quoted as 50%. That is true for prompts up to 272,000 input tokens. Past that threshold, OpenAI applies its long-context rates to the whole request — not just the portion above the line. Anthropic serves Opus 5.5's full 1M-token window at flat rates.

The difference is not gradual. Here is a request with 20,000 output tokens at both sides of the threshold:

That threshold turns ordinary agent decisions into cost decisions. An agent that keeps appending full documents to its context, or returns large tool results without compaction, will cross 272K without anyone intending it to. But the reverse is also true, and worth saying: compaction and retrieval discipline are now worth real money on Sol specifically, because staying under the line keeps the 50% advantage intact.

Where the half-price headline holds, and where it evaporates

The pattern is clean: short prompts and output-heavy work keep the full 50%. Large fresh document inputs with little output — the shape of RAG pipelines, contract review and compliance screening — collapse it to between 1% and 5%. If your workload looks like the last two rows, the price is not the reason to switch.

The caching contracts are not the same product. Sol requires a 1,024-token minimum and retains a cache for at least 30 minutes from the last write or reuse, refreshing on reuse. Opus 5.5's minimum is 512 tokens with a five-minute default or a pricier one-hour option. Sol's cheaper cache reads are real, but a workflow that goes quiet for an hour, or whose prefix shifts between calls, may find Opus's longer retention cheaper in practice despite the higher rate.

Speed: the counterintuitive part

Both vendors market throughput, and both are right about their own number. What the independent run shows is that throughput and task time are different quantities, and they favour different models.

Opus 5.5 streams faster at every setting — 97 tokens per second at max against Sol's 58 — which is why Anthropic's "more than 30% faster output" claim holds. But it takes 12.4 minutes to produce its first token at max effort, against 4.7 minutes for Sol. End to end, Sol finishes an index task in 647 seconds where Opus takes 783.

That has a practical consequence for interactive use. A model's streaming speed is what you feel while reading an answer; its time-to-first-token is what you feel while waiting for one to start. For a chat interface, Opus 5.5's 5.4-second-equivalent settings matter — at medium it starts in 23.5 seconds against Sol's 5.4. For a batch agent that runs unattended, only the total time per task counts, and Sol wins there at every comparable price point.

The same model, three different scores

The most instructive part of this comparison is how much Opus 5.5's numbers move depending on who ran them. These are not errors; they are different harnesses, and they should put a floor under how much weight you give any single figure.

Three things follow. First, AutomationBench is two benchmarks sharing a name: Zapier's binary workflow score and Artificial Analysis's partial-credit version read 40.0% and 69.5% for the same model at the same effort. Quoting either without the qualifier is misleading. Second, OpenAI's GDP.pdf figure for Opus 5.5 excludes nothing but is labelled "with fallbacks", and OpenAI itself notes that its AutomationBench figure for Claude Fable 5.1 omits fallback cost that occurred on roughly 40% of tasks. Third, on Anthropic's side, the Opus 5.5 launch table has no row for GPT-6.1 Sol — it lists GPT-5.6 Sol, the model OpenAI replaced eight days before.

Neither vendor is being dishonest. Both are showing where their model looks best. The reason the Artificial Analysis table anchors this article is simply that it is the only run where both models faced the same tests under the same operator.

Safety and safeguards: two different bets

The fallback policy changes what the benchmarks measure. When Opus 5.5's classifiers fire, work is completed by a different model — and Vals records how often that happened on its index: 91 of 2,281 tasks (3.99%) were fallback-assisted, with 18 refusals (0.79%) scored as failures. On SRE-Bench, outside the index, 82.82% of Opus's tasks were fallback-assisted and 10.30% were refusals, which is why Opus scores 33.59% there against Sol's 50.76%. That is not Opus 5.5 doing security work badly; it is Opus 5.5 declining to do it and handing it to Opus 4.8.

Read positively, that is the safeguard working. Read as a benchmark, it means a "Claude Opus 5.5" row can be a mixture of three models. For security and biology workloads specifically, check your refusal path before comparing scores.

Operational changes to audit before you switch

• effort: "none" is gone. GPT-6.1 Sol drops the non-reasoning setting that GPT-6 Sol offered. OpenAI's migration guide says to use low instead, so any caller setting none explicitly needs a config change and a regression run.
• Sol uses 10–30% more output tokens than GPT-6 Sol at each effort setting, per Artificial Analysis. The cost cut came from the rate card, not from efficiency.
• Sol is slower to stream: 58–59 tokens per second at its top settings against GPT-6 Sol's higher figures and an average of about 70 across tracked models. Latency-sensitive interfaces should be tested before migrating.
• Opus 5.5 cannot disable thinking at all. If your integration shuts thinking off to control cost, it now returns a 400 — low effort is the replacement.
• Refusal billing changed on 24 September 2026. Anthropic resumed charging for pre-output refusals in the bio, frontier_llm and reasoning_extraction categories; cyber refusals remain unbilled. If your pipeline triggers safeguards, that is a new line on the invoice — and a fallback is charged separately at the fallback model's rates.
• Cache minimums differ: 512 tokens for Opus 5.5, 1,024 for Sol. Short system prompts that cache on one model may not cache on the other.
• Availability: Sol is in the API, ChatGPT Work, Codex, Bedrock and OpenRouter but not in regular ChatGPT chat, and it is off by default in Enterprise and Edu workspaces until an admin enables it. Opus 5.5 is on five cloud platforms from day one.
• One dead end to name: "cost per accepted task" is the right production metric, and neither vendor publishes it. If Sol attempts a task at $0.40 and passes 40% of the time while Opus attempts at $0.80 and passes 90%, the arithmetic is close to a wash despite a 2× sticker gap. Measure success rates on your own queue.

Verdict

FAQ

Sources & further reading

• Artificial Analysis — GPT-6.1 Sol vs Claude Opus 5.5 release comparison: all ten effort levels, index scores, cost per task, component evaluations, output speed and time per task.
• Artificial Analysis — GPT-6.1 Sol replaces GPT-6 Sol after just 7 days: the independent launch analysis, cost-efficiency frontier and token-usage findings.
• OpenAI — Introducing GPT-6.1 Sol: DeepSWE v1.1, OSWorld 2.0, GDP.pdf, AutomationBench, Terminal-Bench Science and factuality charts.
• OpenAI — GPT-6.1 Sol system card addendum: capability classifications, safety evaluations, broken-tool disclosure and guardrail adherence.
• Anthropic — Introducing Claude Opus 5.5 and the Opus 5.5 model documentation: pricing, benchmark table, context and output limits.
• Claude Platform Docs — Refusals and fallback and the release notes (24 September 2026): refusal billing categories and fallback credit.
• Vals — Index v2.1 with the GPT-6.1 Sol and Opus 5.5 model pages: the second independent composite, plus fallback and refusal accounting.
• Kingy AI — GPT-6.1 Sol vs Claude Opus 5.5: benchmarks, specs and cost per task: the reproducible per-request cost table and the 272K threshold arithmetic.
• Vellum — GPT-6.1 Sol benchmarks explained: chart-data tables for DeepSWE, OSWorld 2.0, GDP.pdf and AutomationBench.
• DataCamp — GPT-6.1 Sol: features, benchmarks, pricing and access.
• Related on CodingFleet: Claude Opus 5.5 Review, Claude Opus 5.5 vs GPT-6 Sol, GPT-6 Sol vs GPT-5.6 Sol, GPT-6 Astra Review.

Benchmark scores are vendor- or evaluator-reported and labelled as such throughout. Effort settings are stated wherever the source states them; where a vendor omits the effort level of a comparison, this article says so rather than assuming one. Cross-harness numbers are directional. Prices are USD per million tokens as of early October 2026.