Claude Opus 5 tops FrontierBench at 43.3% while holding $5 input pricing

Analyses published over the weekend position Claude Opus 5, which Anthropic launched Friday, as the price performance leader among frontier agentic models. Anthropic's launch materials list standard pricing of $5 per million input tokens and $25 per million output tokens, a fast mode at $10 and $50, a 1 million token context window and 128,000 tokens of maximum output, alongside a 43.3 percent score at maximum effort on FrontierBench v0.1, the 74 task successor to Terminal-Bench. Weekend reviews from BuildFastWithAI and Codersera both put the model ahead of OpenAI's GPT-5.6 Sol on that benchmark, though they disagree on the size of the gap.

The numbers Anthropic published

All of the benchmark figures circulating this weekend originate in Anthropic's own launch materials, a caveat worth keeping in view until independent replications land. Those materials report 43.3 percent at maximum effort and a 44.4 percent mean reward at the xhigh effort setting on FrontierBench v0.1. Codersera notes the result more than doubles what Opus 4.8 posted on the same evaluation, at a lower cost per completed task. BuildFastWithAI notes the standard pricing matches Opus 4.8, meaning the capability gain arrives with no per token price increase, which is the core of the price performance argument. The 1 million token context window applies as both default and maximum, per Codersera, and the 128,000 token output ceiling is a practical detail for agent builders whose workloads generate long tool call transcripts.

The Sol gap depends on who is counting

The comparison with GPT-5.6 Sol is murkier. Codersera puts Sol at 34.4 percent on FrontierBench, while BuildFastWithAI cites 37.5 percent. On those numbers, Sol trails Opus 5 by roughly six to nine points depending on the analysis, and neither outlet explains the discrepancy, which may reflect different effort settings or benchmark revisions. What the weekend analyses agree on is the direction: Opus 5 currently leads the benchmark among frontier agentic models, per the figures in circulation.

Four flagships in under two months

BuildFastWithAI counts four flagship releases from Anthropic in under two months: Mythos 5, Fable 5, Sonnet 5 and now Opus 5. That cadence is itself a competitive statement, and it sharpens the question of whether safety evaluation can keep pace with release velocity; the model's system card, including results from UK government testing, is examined in our companion report on the Opus 5 system card. We assessed the launch economics in our opinion column on Opus 5, and much of the specification matched what surfaced in the pre-launch leak we reported July 14.

Where it runs today

Distribution moved quickly. Opus 5 has been available in GitHub Copilot for Pro+, Max, Business and Enterprise plans, and on Amazon Bedrock, since Friday, which puts the model in front of enterprise developers without any additional procurement step. Neither company has published head to head production comparisons, and OpenAI has not responded publicly to the weekend analyses. For agentic workloads, that availability compounds the pricing story: agents consume output tokens heavily and rerun tasks until they succeed, so cost per completed task, not cost per token, is the number buyers ultimately pay. The weekend analyses converge on the claim that Opus 5 currently offers the best ratio of benchmark performance to token price among frontier models. That claim rests on vendor published scores and two independent writeups that do not fully agree with each other, so it should be read as the opening position in a pricing argument, not the settled outcome. The rebuttals, from OpenAI and from independent benchmarkers, will arrive within days.