Anthropic's newest mid-tier model has claimed third place on Arena.ai's Agent Arena leaderboard. Claude Sonnet 5.5, which launched on September 28, 2026, posted a net improvement score of 12.5% — enough to edge out OpenAI's top entry. Only two models rank above it, and both are also made by Anthropic.

How the leaderboard shakes out

Within days of its release, Sonnet 5.5 had climbed to the number three spot on Agent Arena. Its precise net improvement score is 12.52%, with a margin of error of plus or minus 3.09%.

The two models ahead of it are Claude Fable 5.1 (Max) at 14.31% and Claude Opus 5.5 (High) at 13.82%. Directly below sits OpenAI's GPT-6 Astra (Max) at 12.27%. With those results, Anthropic holds three of the top four positions on the board.

Agent Arena grades models on millions of real-world agentic tasks, in which an AI has to use tools, complete multi-step jobs, and satisfy actual users. Scoring incorporates tool reliability, task completion, and user feedback. That emphasis tracks a broader industry shift: as deployments lean on models to call tools and carry out multi-step work, model rankings are increasingly contested on end-to-end task execution rather than static question-answering alone.

Cost is the other headline number. Arena lists Sonnet 5.5 at $2.74 per task, while early October snapshots place its median cost at roughly $2.78 per task. The model also consumes more tokens than most of its peers.

The benchmarks beyond Arena

Sonnet 5.5's strong showing extends to other benchmarks as well. On the Artificial Analysis Intelligence Index, it scored 56, good for second place behind Opus 5.5 (Max).

On Terminal-Bench 4.0, Sonnet 5.5 scored 70.6% — a substantial jump from its predecessor, Sonnet 5, which managed roughly 10 to 14% on the same test. The new model also edged out Opus 5.5, which scored 66.4% on Terminal-Bench 4.0.

Speed improved as well: Sonnet 5.5 runs more than 30% faster than Sonnet 5. Pricing, however, did not move. The model still costs $2 per million input tokens and $10 per million output tokens, the same as before. Sonnet 5.5 offers a context window of 1 million tokens and a maximum output of 128,000 tokens. With per-token prices held flat while measured speed and scores rose, the upgrade shifts the capability-per-dollar equation — though the model's heavier token appetite, visible in its Arena figures, remains the variable that determines real-world cost per job.

What this means for the AI model race

The first thing to watch is the error bar. Sonnet 5.5's score carries a margin of plus or minus 3.09%, and the overall spread separating the top four models is smaller than that. GPT-6 Astra (Max) trails Sonnet 5.5 by just 0.25 percentage points. With gaps that narrow, the ordering can change as Arena aggregates more task results, so the standings are best read as a current snapshot rather than a settled hierarchy.

Token consumption matters here too. Higher token usage can eat into a cost advantage depending on the workload — a model that is cheap per token but verbose may not always be cheap per job. For teams running agents in production, those two columns — score and cost per task — end up on the same spreadsheet, which is why any release that reshuffles the top of boards like Agent Arena gets weighed on both dimensions.