OpenAI's GPT-6.1 Sol and Claude Sonnet 5.5 match on price, launch a day apart, share no benchmarks

Same sticker price, one day apart

Anthropic released Claude Sonnet 5.5 on 28 September 2026, and OpenAI followed with GPT-6.1 Sol on 29 September. According to a comparison published on dev.to, both models list at $2 per million input tokens and $10 per million output tokens, with matching batch rates of $1/$5. Despite the head-to-head timing, neither company has compared the two directly: OpenAI benchmarked GPT-6.1 Sol against GPT-6 Sol, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1, while Anthropic measured Sonnet 5.5 against Sonnet 5, Opus 5.5 and GPT-6 Sol — the previous OpenAI model, since GPT-6.1 Sol did not exist when Anthropic ran its numbers.

Where the bills actually diverge

Identical list prices do not mean identical invoices. The dev.to write-up highlights several splits:

• Cache reads cost $0.10 per million tokens on GPT-6.1 Sol versus $0.20 on Sonnet 5.5, which favors OpenAI for workloads that reuse long system prompts. Cache writes are $2.50 on GPT-6.1 Sol; Anthropic charges $2.50 for a five-minute cache and $4 for one hour.
• Prompts above 272,000 input tokens trigger higher rates on the entire GPT-6.1 Sol request — $4 input, $15 output — while Sonnet 5.5 keeps standard pricing across its one-million-token window. A 400,000-token prompt with a 5,000-token response costs roughly $1.68 on GPT-6.1 Sol and $0.85 on Sonnet 5.5 without caching.
• OpenAI sells a paid Fast tier at $4/$20, with an "Ultrafast" tier announced; Sonnet 5.5 has no speed tier.
• Context windows are near parity — 1,050,000 for GPT-6.1 Sol with a 922,000 input cap, versus 1M for Sonnet — and both cap output at 128,000 tokens, though Sonnet reaches 300K in batch with a beta header. Sonnet's knowledge cutoff is June 2026, later than GPT-6.1 Sol's 30 April 2026.
• Distribution differs: GPT-6.1 Sol is on OpenRouter, while Sonnet 5.5 ships on Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS.

Competing claims with mismatched setups

OpenAI pitches GPT-6.1 Sol as delivering "near-Astra intelligence" at one-fifth of Astra's standard token price. Its launch post cites AutomationBench 1.0.6 at medium effort (+2.2 points over Opus 5.5 at roughly a third of the cost, +4.8 over GPT-6 Sol), Terminal-Bench Science 0.1 at max effort ($5.47 per task versus $23.21 for Opus 5.5, with GPT-6 Astra still tops at 68.1%), and OSWorld 2.0 offline (+7 points over GPT-6 Sol for under half the cost).

Anthropic's headline claims over 30% more speed than Sonnet 5 and up to 30% lower cost per task. On FrontierCode 1.1, run by Cognition, Sonnet 5.5 scored 52.1% at xhigh and 46.2% at max against 49.3% for GPT-6 Sol, and at high effort Anthropic says it matches GPT-6 Sol's best score at about one-fifth the cost. On Artificial Analysis-run GDPval-AA v2.1 and AA-Briefcase v1.1, Sonnet 5.5 scored 1844 versus 1487 and 1811 versus 1483.

Per dev.to, these figures cannot be merged into one table: OpenAI ran AutomationBench at medium effort in its own harness while Anthropic's summary tables use max effort for Claude models; Anthropic reports 59.9% for Sonnet 5.5 on Terminal-Bench Science where OpenAI published no raw GPT-6.1 Sol score; and OpenAI used OSWorld 2.0 offline while Anthropic used OSWorld 2.1.

Third parties still test the old OpenAI model

The most complete independent comparison, Artificial Analysis's Intelligence Index v4.3.2 aggregating ten evaluations, pairs Sonnet 5.5 with GPT-6 Sol, not GPT-6.1. Sonnet scores higher at every effort level but costs more per task at the same level: from index 41 at $0.59 per task on medium effort up to 56 at $7.60 on max, versus GPT-6 Sol's 40 at $0.25 to 48 at $1.05. At max effort Sonnet generates 138 output tokens per second against 76 for GPT-6 Sol. A fairer cross-effort pairing, per dev.to, puts Sonnet 5.5 at high effort (index 47, $1.08) against GPT-6 Sol at max (index 48, $1.05). OpenAI claims GPT-6.1 Sol beats GPT-6 Sol at equal or lower effort, so these numbers may shift once it is independently measured.

Why it matters

Two frontier models at the same list price released a day apart invites a direct comparison that no vendor has actually run. The gap between price per token and cost per task — driven by token consumption, differing default effort levels, caching and long-prompt surcharges — means the cheaper model depends entirely on the workload. Until evaluators test GPT-6.1 Sol, the only reliable way to choose, as dev.to concludes, is to run both on your own tasks and measure cost per successful task rather than per token.