BuildNot yet confirmed elsewhere1 publisher2 min readPublished
Claude Sonnet 5.5 outscores Opus 5.5 on Terminal-Bench 4.0, according to a dev.to post
Anthropic's Claude Sonnet 5.5 resolves 70.6% of Terminal-Bench 4.0 tasks against 66.4% for Opus 5.5, a dev.to post reports. If that gap holds on a team's own tasks, the mid-tier model is the one to test first as the default for terminal agents.
The Engineer · Build desk
Bar comparison of Terminal-Bench 4.0 resolution rates: Claude Sonnet 5.5 at 70.6%, Claude Opus 5.5 at 66.4%, and GPT-6 Astra at 61.8%, per a dev.to post.
Terminal-Bench 4.0 resolution rate, per a dev.to post
| Item | Value | Claim |
|---|---|---|
| Claude Sonnet 5.5 | 70.6% | 6 |
| Claude Opus 5.5 | 66.4% | 7 |
| GPT-6 Astra | 61.8% | 8 |
What happened
- The post places Sonnet 5.5 between Haiku 5.5 and Opus 5.5 in Anthropic's lineup, as the company's main model for corporate work.
- OpenAI's GPT-6 Astra scored 61.8% on the same benchmark, according to the post.
- The post says the model was tuned for terminal work, including bash, stdout and stderr redirection, process signals and the GDB and LLDB debuggers.
- The post lists generation 30% faster than intermediate generations, a 128,000-token output ceiling and a 1,000,000-token context window.
Why it matters
- decision Teams that default to Opus 5.5 for shell agents now have a benchmark reason to trial Sonnet 5.5 against it before their next model review.
- exposure Until Anthropic or the benchmark's maintainers publish matching results, any default switched on the 70.6% figure rests on one community post.
- cost The post ties the efficiency to aggressive prompt caching, so the savings go mostly to pipelines that resend the same long context on every turn.
In an agent loop, every failed task means a retry or a handoff to a person, so the failure rate is the figure to price. At 70.6%, Sonnet 5.5 misses 29.4% of Terminal-Bench 4.0 tasks. Opus 5.5 misses 33.6% [2]. On those figures the mid-tier model fails 12.5% fewer tasks [3].
The benchmark, the post says, is end-to-end repair. The model has to diagnose a bug in a real repository, install missing dependencies, run the tests, read the stack traces, apply the patch and confirm the build, all without human intervention [9]. For the 4.2-point gap [1] to carry over, a team's agent tasks have to look like that. Its harness also has to give each model the same tools, turn limits and retry policy the benchmark runs used. The two scores also have to come from comparable runs.
The post sells Sonnet 5.5 on capability as much as on efficiency. On its own table the middle tier beat the top one [1]. It also says the model beat Google DeepMind's Gemini 4 Argon and set the state of the art for Linux shell work, CI pipeline debugging and autonomous repository management [16].
The efficiency argument covers the whole category. According to the post, inference at 12 to 20 per million input tokens kept frontier models on one-off tasks through the first half of 2026 [11]. High latency added to the problem, and background loops running hundreds of iterations per hour were out of reach [11]. The post did not disclose who ran the benchmark, on what scaffold, Sonnet 5.5's own price, what currency that range is in, or a Haiku 5.5 score.
We think the gap is large enough to justify running both models side by side on a team's own backlog of failed agent tasks before the default changes. The rest of the post's evidence is further from pipeline work. It reports 61.6% on the Chartography benchmark without external OCR tools [12]. It also puts the model through Pokemon Red, played only from Game Boy screenshots [15], a workload few CI pipelines share.
What to watch
- Anthropic's own model card or price list for Sonnet 5.5, which would confirm or revise the 70.6% figure and put a per-token price on it.
- An independent Terminal-Bench 4.0 run with a published harness that puts Sonnet 5.5 and Opus 5.5 side by side.
- A Haiku 5.5 score on the same benchmark, showing whether the cheapest tier is close enough for high-volume agent loops.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence20
- Adoption
- Insufficient
- Hype gap+45
- Incentives
- Insufficient
- Confidence25
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Sonnet 5.5 leads Opus 5.5 on Terminal-Bench 4.0 by 4.2 percentage points.
- [2]
Sonnet 5.5 fails 29.4% of Terminal-Bench 4.0 tasks and Opus 5.5 fails 33.6%.
- [3]
Sonnet 5.5 fails 12.5% fewer Terminal-Bench 4.0 tasks than Opus 5.5, relative to Opus's failure rate.
- [4]
According to the dev.to post, Anthropic officially launched Claude Sonnet 5.5, and the model is catalogued and tracked on Models.dev.
- [5]
The post says Claude Sonnet 5.5 is built as Anthropic's main model for corporate work and sits between the lightweight Haiku 5.5 and the deep-reasoning Opus 5.5.
- [6]
Claude Sonnet 5.5 scored a 70.6% resolution rate on Terminal-Bench 4.0, according to the post.
- [7]
Claude Opus 5.5 scored 66.4% on Terminal-Bench 4.0, according to the post.
- [8]
OpenAI's GPT-6 Astra scored 61.8% on Terminal-Bench 4.0, according to the post.
- [9]
The post describes Terminal-Bench 4.0 as testing whether a model can diagnose a bug in a real repository, install missing dependencies, run tests, analyse stack traces, apply the correct patch and validate the final build without human intervention.
- [10]
The post says Sonnet 5.5 offers generation speed 30% higher than intermediate generations, a 128,000-token output ceiling and a 1,000,000-token context window.
- [11]
The post says recurring inference costs of 12 to 20 per million input tokens and high latency limited frontier models to one-off tasks during the first half of 2026, ruling out background execution loops with hundreds of iterations per hour.
- [12]
Sonnet 5.5 scored 61.6% on the Chartography benchmark without external OCR tools or helper scripts, according to the post.
- [13]
The post ties the efficiency of agents like Sonnet 5.5 to support for aggressive prompt caching and persistent reasoning it calls Preserved Thinking.
- [14]
The post says Sonnet 5.5 was tuned for terminal operation, including bash execution, stdout and stderr redirection, process signal handling, the GDB and LLDB debuggers, and navigation of very large file trees.
- [15]
The post says the model was set to play the Game Boy game Pokemon Red operating only from screenshots.
- [16]
The post claims Sonnet 5.5 surpassed Claude Opus 5.5 and more expensive frontier peers including OpenAI's GPT-6 Astra and Google DeepMind's Gemini 4 Argon, setting the state of the art in Linux shell interaction, CI pipeline debugging and autonomous repository management.
Sources
1 independent publisher whose own reporting we read for this story.
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.