Anthropic’s Smaller Model Sonnet 5.5 Overtakes OpenAI’s Flagship GPT-6
Set Trending Topics as a preferred source on Google.
A midrange model that beats OpenAI’s best: Anthropic has just released Claude Sonnet 5.5, and it lands straight at No. 2 in the most closely watched independent A.I. ranking. On the Intelligence Index from Artificial Analysis, Sonnet 5.5 scores 56 points, just two behind Anthropic’s own flagship Opus 5.5, which took the top spot only days ago. That gives Anthropic the first two places. OpenAI’s GPT-6 Astra trails with 53 points, and GPT-6 Sol, its direct rival on price, with 48.
An 18-Point Jump in One Generation
The leap over its predecessor is substantial: Sonnet 5 scored 38 points on the same index, so Sonnet 5.5 gains 18. Artificial Analysis runs every model through the same set of benchmarks, from coding tasks and knowledge questions to agents that work on a computer on their own. That makes the comparison more neutral than the figures vendors publish themselves.
The current Intelligence Index ranking:
- Claude Opus 5.5 (Anthropic): 58 points
- Claude Sonnet 5.5 (Anthropic): 56 points
- GPT-6 Astra (OpenAI): 53 points
- GPT-6 Sol (OpenAI): 48 points
- Claude Sonnet 5 (Anthropic): 38 points
- GPT-6 Luna (OpenAI): 37 points
On Par With Opus for Agents and Office Work
Sonnet 5.5 is especially strong where A.I. agents work independently. On Terminal-Bench 4.0, which measures how well a model completes tasks via the command line, Sonnet 5.5 reaches 64 percent, according to Artificial Analysis. That is ahead of Opus 5.5 and GPT-6 Astra, both at 60 percent, and a 50-point gain over Sonnet 5. On Terminal-Bench-Science, where agents run through realistic research workflows, Sonnet 5.5 scores 53 percent, behind only GPT-6 Astra and Opus 5.5.
In classic knowledge work, such as building presentations, analyses and documents, Sonnet 5.5 essentially matches Opus 5.5:
- GDPval-AA: 1,844 vs. 1,846 Elo points
- AA-Briefcase: 1,811 vs. 1,822 Elo points
- AutomationBench-AA: 71 vs. 70 percent
The smaller model is weaker on factual knowledge. On AA-Omniscience, Sonnet 5.5 answers 54 percent of questions correctly, compared with 66 percent for Opus 5.5. It hallucinates less often, though, at a rate of 47 percent versus 59 percent for Opus. On Humanity’s Last Exam and scientific coding (SciCode), Sonnet also sits about six points behind its bigger sibling.
The Catch: More Tokens Than Any Other Model
That top performance comes at a cost. To reach Opus-level results in its highest reasoning mode, Sonnet 5.5 generates about 193,000 output tokens per task, according to Artificial Analysis. That is the highest figure the testers have ever measured, roughly 60 percent more than Opus 5.5 and about seven times as much as GPT-6 Astra.
Anthropic left the list price unchanged. Like Sonnet 5, Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens (about 1.70 and 8.50 euros). That is exactly the price of GPT-6 Sol, which OpenAI recently launched at half the price of the previous generation. Because of its heavy token use, however, a single Intelligence Index task costs about $7.60 with Sonnet 5.5, roughly 50 percent more than with Sonnet 5. For comparison: Opus 5.5 comes in at $5.98, GPT-6 Astra at $3.26 and GPT-6 Sol at $1.06.
Anthropic itself promises that Sonnet 5.5 costs up to 30 percent less than its predecessor “for most work” because it operates more efficiently. That refers to everyday use, not the maximum mode. At lower reasoning levels, Artificial Analysis sees a mixed picture: the “high” setting is the most competitive, landing just behind GPT-6 Sol on intelligence at effectively the same cost per task. At the lowest settings, GPT-6 Sol delivers more performance with fewer tokens.
Artificial Analysis notes one caveat itself: it tested a pre-release version with a bug affecting structured outputs, which Anthropic fixed for the public launch. Results should barely change, but some tests will be re-run.
Luna Still Has No Challenger
Sonnet 5.5 is available now in the Claude apps, via the Anthropic API and on Amazon Web Services, Google Cloud and Microsoft Azure. Several companies tested the model ahead of launch. Sonnet 5.5 “cleared the same quality bar you’d expect from a higher-tier model,” says Daniel Vogel, COO of Epic Games. Also new: for the first time, Anthropic is using safeguards against distillation on a Sonnet model, meaning attempts to extract the model’s reasoning in order to train other models on it.
One gap remains. With GPT-6 Luna, OpenAI offers a model at $0.10 per million input tokens, one-twentieth of Sonnet 5.5’s price. Luna scores only 37 points on the Intelligence Index, but a task costs just 7 cents. Anthropic’s answer, Claude Haiku 5.5, is due in the coming weeks.
Competition remains fierce. Google keeps expanding its Gemini models, Meta has started selling its Muse models to businesses through the new Meta Enterprise Platform, and Chinese providers like DeepSeek undercut U.S. prices significantly with open-weight models. At the top of the ranking, though, there is only one name for now: Anthropic.
