Anthropic releases Claude Sonnet 5.5 with 30% speed and cost gains

By Splitfeed EditorialLast updated Sep 29, 2026, 10:30 UTC

On September 28, 2026, Anthropic released Claude Sonnet 5.5, the second model in its Claude 5.5 family. The company says the model generates output more than 30 percent faster and costs up to 30 percent less per task compared to prior versions. Its most striking benchmark result comes from Terminal-Bench, a coding evaluation, where it scored 70.6 percent against a previous score of 10.3 percent. Anthropic also reports that Sonnet 5.5 nearly matches the more expensive Claude Opus 5.5 on knowledge-work benchmarks, suggesting significant performance compression across the model tier. A third model, Haiku 5.5, is announced for release in the coming weeks.

The broader context gives this release weight beyond a routine update. Once Haiku 5.5 ships, Anthropic will have a direct counterpart to each of OpenAI's three GPT-6 models, completing a competitive lineup at the budget, mid-tier, and flagship levels. The release is also happening as Anthropic moves toward an IPO, meaning each model launch carries commercial and investor significance alongside its technical one. The Terminal-Bench jump, from 10.3 to 70.6 percent, is large enough that it will draw scrutiny over what that benchmark actually measures and how representative it is of real-world coding tasks.

None of the three camps, Pro-AI, Anti-AI, or Middle Ground, have published reactions to this story yet. Pro-AI voices would typically treat a simultaneous speed gain, cost reduction, and benchmark improvement as validation that AI capabilities are scaling efficiently and becoming more accessible. Anti-AI commentators would likely question whether benchmark scores translate to reliable real-world performance and raise concerns about the pace of deployment outrunning safety evaluation. Middle Ground observers would probably focus on the competitive dynamics, noting that faster and cheaper models lower the barrier to deployment while acknowledging that independent verification of the claimed gains is still needed.

The clearest near-term signal to watch is the release of Haiku 5.5 and any independent evaluations of Sonnet 5.5 on tasks outside Terminal-Bench. Third-party assessments of whether the knowledge-work parity with Opus 5.5 holds in practice would go a long way toward settling the question of how much the benchmark numbers reflect genuine capability gains.