Claude Sonnet 5.5 Is Half the Price of Opus 5.5. On Terminal-Bench 4.0, It Costs 25% More.

The leaderboard

Vals.ai runs Terminal-Bench 4.0: 66 new, community-contributed, maintainer-reviewed tasks across software, science, ML, operations, hardware, security and media, none of them reused from Terminal-Bench 2.1. Its leaderboard was last updated October 24, 2026. The summary line reads: Claude Opus 5.5 “ranks first on Terminal-Bench 4.0 with 65.15%, followed by Claude Sonnet 5.5 (Mini-SWE-agent) at 64.14% and Kimi K3.1 Pro (Mini-SWE-agent) at 63.64%.”

The ranking is the headline everyone will repeat. The cost column next to it is the story. Here are the top 8, plus the 2 cheapest models worth a look.

Source: Vals.ai Terminal-Bench 4.0 leaderboard, updated October 24, 2026. Every entry uses the mini-swe-agent harness. Cost, price and duration columns as published by Vals.ai.

How half the price ends up costing more

Anthropic’s pricing page makes the math clean. Every Sonnet 5.5 rate is exactly half the matching Opus 5.5 rate: $2 vs $4 input, $10 vs $20 output, $0.10 vs $0.20 for cache hits, $2.50 vs $5 for 5-minute cache writes, $4 vs $8 for 1-hour cache writes.

So the token mix doesn’t matter. If Sonnet 5.5 bills 25% more at half the rate on every line, it used roughly 2.5 times the price-weighted token volume of Opus 5.5 on the same 66 tasks. That’s our arithmetic on Vals.ai’s published numbers, assuming Vals.ai bills at the list rates it shows, not a token count Vals.ai reports.

The duration column points the same way. Sonnet 5.5 took 1h22m, Opus 5.5 took 1h04m, 18 minutes longer. A cheaper model that needs more steps to reach the same place hands the saving back, with interest.

Terminal-Bench 4.0 is built to expose exactly this. Vals.ai describes its tasks as “long, multi-step, and graded strictly on the final artifact.” The harness gives each model one bash tool and “no step or cost limit; the agent works until it signals that it has finished or reaches the task’s time limit,” which is 8 hours. Nothing caps a model that wanders.

The gap is smaller than one task

Terminal-Bench 4.0 has 66 tasks, so one task is worth 1.52 points. Each score is the mean of 3 runs, with no partial credit: the full verifier suite must pass.

Opus 5.5 leads Sonnet 5.5 by 1.01 points, about two-thirds of a task per run. Kimi K3.1 Pro trails Opus 5.5 by 1.51 points, about 1 task. Vals.ai publishes standard-error bars for each score; on a 66-task test, gaps this small are a tie in practice.

That makes the cost column decisive. If the top 3 are within a task of each other, the question becomes what each one costs to get there.

Cost per point = Vals.ai cost per test ÷ Vals.ai score, rounded to the cent. A rough efficiency read, not a Vals.ai metric.

4 more numbers from the same table

Kimi K3.1 Pro costs 65% less than Opus 5.5 and 72% less than Sonnet 5.5. Vals.ai lists it at the same $2 / $10 token price as Sonnet 5.5. Same rate, $4.57 against $16.51 per test, 0.5 points apart. Token efficiency, not list price, separates them.

Claude Fable 5.1 is Anthropic’s most expensive model and scores below both of its cheaper siblings. At $10 / $50, 2.5 times Opus 5.5’s rates, it scores 58.08% at $17.18 per test, 7.07 points under Opus 5.5 and 30% more expensive.

Opus 5.5 is both better and cheaper than Opus 5. Opus 5 scores 53.53% at $18.60 per test. Opus 5.5 adds 11.62 points and costs 29% less per test. Its list price also dropped, from $5 / $25 to $4 / $20. The older Sonnet 5 scores 10.61% at $26.05 per test, so Sonnet 5.5 is a far bigger generational jump than its price row suggests.

Haiku 5.5 is 40 times cheaper per token than Opus 5.5, and 4.2 times cheaper per test. Its $0.10 / $0.50 rate applies to prompts up to 100,000 tokens; above that, Anthropic charges $0.50 / $2.50. It scores 36.36%, so it fits sub-tasks, not the whole agent loop.

What this means if you code with AI

Stop picking agent models off the per-token price. For chat, short edits and autocomplete, token price tracks cost well. For agents that loop until they decide they’re done, the model’s token appetite matters as much as its rate. Sonnet 5.5 is the proof: half the rate, a bigger bill.

If you pay Anthropic per token for agent runs, test Opus 5.5 before defaulting to Sonnet 5.5. On this benchmark it’s faster, slightly more accurate and 20% cheaper per test. Your codebase is not Terminal-Bench, so measure cost per finished task on your own work for a week before switching.

If you’re on a flat subscription, the math changes. On a subscription plan you don’t see a per-token bill, you see usage limits. A model that burns 2.5 times the tokens will hit those limits sooner, whatever its rate card says.

Put cheaper models on the shortlist. Kimi K3.1 Pro lands within about 1 task of the leader at roughly a third of its cost. GPT-6.1 Sol gives up 10.10 points for an 87% lower cost per test. For many teams, routing the bulk of agent work to a model like that and saving the frontier model for the hard tickets is the cheaper setup.

Set a spend cap on any agent with no step limit. Terminal-Bench 4.0 lets models run for up to 8 hours per task. Your own tools may not stop sooner unless you tell them to.

Where MegaBrain fits

Full disclosure: we sell an AI coding agent. MegaBrain’s terminal coding agent runs GLM, Kimi, MiniMax, DeepSeek, Qwen, Grok and Auto on the paid plans, Claude, GPT and Gemini from credits from Pro, and any model through your own key or local models, free on every plan. So you can run the same task on Opus 5.5 and a Kimi model and compare the bills yourself. Plans are on the pricing page. MegaBrain is not affiliated with Vals.ai or Anthropic.

Sources

All pages opened October 26, 2026.

• Vals.ai — Terminal-Bench 4.0 leaderboard and methodology (scores, cost per test, duration; updated October 24, 2026)
• Anthropic — Claude API pricing (Opus 5.5, Sonnet 5.5, Haiku 5.5, Fable 5.1 per-token and cache rates)