Claude Sonnet 5.5 vs Opus 5.5: Terminal Bench 4.0 and the Cost Per Task Math

Anthropic shipped Claude Sonnet 5.5 on September 28, 2026, and published its own head to head against Opus 5.5 rather than hiding the comparison in a footnote. That is worth noticing on its own: vendors usually frame the cheaper model as good enough. Anthropic's own benchmark table instead shows Sonnet 5.5 outright winning one of the harder agentic benchmarks while trailing on five others, all at half the per token price. That is a more useful signal than another "nearly as good" headline, because it tells you exactly where the fifty percent discount costs you something and where it does not.

The benchmark table, read straight

Benchmark	Sonnet 5.5	Opus 5.5	Who leads
Terminal Bench 4.0	70.6%	66.4% (Xhigh)	Sonnet 5.5
CursorBench 4.0	55.5%	57.8%	Opus 5.5
FrontierCode 1.1	52.1% (Xhigh)	54.4%	Opus 5.5
GDPval AA v2.1	1,844	1,846	Opus 5.5, within noise
OSWorld 2.1 (computer use)	80.1%	81.8%	Opus 5.5
Chartography (no tools)	61.6%	64.4%	Opus 5.5
Input price per 1M tokens	$2	$4	Sonnet 5.5
Output price per 1M tokens	$10	$20	Sonnet 5.5
Context window	1M tokens	1M tokens	Tie

Terminal Bench 4.0 measures real terminal driven agentic tasks: the kind of long running, tool calling, multi step work a coding agent actually does in a shell, not a single function completion. Sonnet 5.5 beating Opus 5.5 there, by more than four points, means the cheaper model is not simply approximating the expensive one. On the other five benchmarks Opus 5.5 leads, but by margins of two to four points, well short of the gap you would expect from a model costing twice as much. Anthropic's own framing calls Opus 5.5 "clearly stronger" on complex, open ended work, and the GDPval and CursorBench numbers support that for ambiguous, judgment heavy tasks. For well scoped terminal and shell driven agent work specifically, the data says reach for the cheaper model first.

What actually changed under the hood

Sonnet 5.5 is not just a price cut on the same model. Anthropic's release notes point to efficiency work that shows up in real deployments rather than only in benchmark tables. Trading firm Balyasny Asset Management reported using about 121,000 tokens per answer on Sonnet 5.5, down from 497,000 tokens on Sonnet 5, a reduction of more than three quarters for comparable tasks. App builder Base44 saw iteration counts drop to 3.6 per app build on Sonnet 5.5, against 7.7 for Opus 5. Zendesk reported tickets processed 20 percent faster. Anthropic attributes the gains to fewer wasted tokens and fewer redundant tool calls, and says output generation itself runs more than 30 percent faster than Sonnet 5. None of that changes the sticker price, listed at the same Sonnet 5 price point of $2 input and $10 output per million tokens, but it does mean the effective cost per completed task can fall well below what the per token price alone suggests, Anthropic cites up to 30 percent lower cost per task against Sonnet 5 from this efficiency work alone, before you even get to the model choice between Sonnet and Opus.

Reasoning also moved. Humanity's Last Exam with tools climbed to 64.5 percent on Sonnet 5.5, up from 54.9 percent on the prior Sonnet generation, a jump large enough that it is not just noise from a harder or easier question set that run.

Where Opus 5.5 still earns the premium

Read the loss column honestly rather than explaining it away. Opus 5.5 wins CursorBench, FrontierCode, OSWorld computer use, and Chartography, and GDPval AA is close enough to call a tie but still tilts to Opus. The common thread across those benchmarks is ambiguity: tasks where the agent has to infer intent, operate a GUI it was not given exact instructions for, or hold a large amount of loosely structured context before producing an answer. Chartography specifically scores answers with no tool access, which isolates raw model reasoning from agentic scaffolding, and Opus 5.5 pulls ahead there by nearly three points. If your workload looks like that, open ended analysis, computer use automation, chart and document reasoning without a deterministic tool to lean on, the premium buys something real.

Context window and max output are identical on paper, 1 million tokens in and 128,000 tokens out on both models, so neither model wins on raw capacity. The difference is entirely in how each model spends that capacity on hard, unstructured problems, which is exactly what the five losing benchmarks measure.

Cross checking against the rest of the field

Anthropic's own comparison table also lists GPT-6 Sol and Gemini 3.1 Pro Preview at similar price points. GPT-6 Sol prices at the same $2 and $10 per million tokens as Sonnet 5.5 but caps favorable pricing at 272,000 input tokens before a multiplier kicks in, and posts a FrontierCode score of 49.3 percent, below both Claude models. Gemini 3.1 Pro Preview prices slightly higher on output at $12 per million tokens and reports neither FrontierCode nor Chartography results, which makes a direct comparison incomplete. On the numbers Anthropic has actually published, Sonnet 5.5 is the strongest terminal agent at its price tier and the cheapest of the three frontier options by a meaningful margin. We walked through a similar open weight versus closed model tradeoff in our Kimi K3 versus Claude Opus 5 comparison, and the pattern holds again here: the gap that matters is not the aggregate score, it is which specific benchmark maps to your actual workload.

A routing rule that follows from the data

Do not pick one model for every agent in your stack. Route terminal driven, tool calling coding agents, CI pipelines, shell automation, deployment scripts, to Sonnet 5.5 first, and measure before escalating. Route open ended research agents, computer use automation, and document or chart heavy analysis to Opus 5.5, where the benchmark gap is real even if it is not large. For a mixed workload, that routing decision alone can cut spend by close to half on the terminal heavy half of the traffic while keeping Opus quality where the data shows it actually matters, which is the same multi tier routing architecture we covered when comparing Opus 5.5 against Gemini 4 Argon and SWE 2.

When this applies to your stack

If you are running a coding agent, a CI assistant, or any tool calling agent against a Claude model today, Sonnet 5.5's Terminal Bench win and half price point make it the default worth testing against your own task set before assuming the more expensive model is required. The benchmark gaps on open ended work are real but modest, which means the decision is rarely all or nothing. Building the routing layer that sends each task to the right tier, and instrumenting it so you can see the actual cost per completed task rather than the sticker price, is exactly the kind of AI integration work that turns a model choice into a unit economics advantage.

FAQ

Is Claude Sonnet 5.5 better than Opus 5.5? It depends on the task. Sonnet 5.5 wins on Terminal Bench 4.0, 70.6 percent against 66.4 percent, which measures real terminal driven agentic work. Opus 5.5 leads on CursorBench, FrontierCode, OSWorld computer use, and Chartography, all tasks that involve more ambiguity or GUI style reasoning.

How much cheaper is Claude Sonnet 5.5? Exactly half the list price: $2 input and $10 output per million tokens against $4 and $20 for Opus 5.5. Anthropic also reports up to 30 percent lower cost per completed task compared to the prior Sonnet 5 generation, driven by fewer tokens and tool calls rather than a price change.

Do Sonnet 5.5 and Opus 5.5 have the same context window? Yes, both support a 1 million token context window and a 128,000 token maximum output. The difference between the models is in reasoning quality on ambiguous tasks, not raw capacity.

Should I route all my coding agent traffic to Sonnet 5.5? Start there for terminal and shell driven agents, since that is where Sonnet 5.5 actually wins the benchmark. Keep Opus 5.5 for open ended analysis, computer use, and document heavy reasoning where the data still favors it, and measure your own escalation rate rather than assuming either model is right for every task.