Claude Sonnet 5 brings stronger agentic AI, plus Anthropic launches Claude Science for researchers | 2026-07-01
Anthropic dropped two releases yesterday: Claude Sonnet 5 and Claude Science. Sonnet 5 is the most agent-capable Sonnet model yet — it can plan tasks, use browsers and terminals, and run through complete workflows on its own. A few months ago, this level of autonomy required much larger and pricier models.
The Sonnet line has a solid reputation among developers. Claude Sonnet 3.5, 3.6, and 3.7 were the first models to show real chops in coding and tool use. More recently, the big agentic gains had been coming from the Opus series. Sonnet 5 narrows that gap considerably — performance is close to Opus 4.8, at lower prices.
Compared to the previous Sonnet 4.6, Sonnet 5 improves across reasoning, tool use, coding, and knowledge work. Anthropic's internal safety evaluations show Sonnet 5 has a lower rate of undesirable behaviors in agentic contexts than Sonnet 4.6, and much weaker cybersecurity capabilities than current Opus models (which is safer for most use cases).
Pricing
Sonnet 5's pricing is aggressive. Introductory rates run through August 31, 2026 — $2 per million input tokens and $10 per million output tokens (about ¥14/¥68). After that, regular pricing kicks in at $3 input and $15 output (¥20/¥102). For reference, Opus 4.8 costs $5/$25 (¥34/¥170).
On BrowseComp (agentic search) and OSWorld-Verified (computer use), Sonnet 5 delivers solid cost efficiency at medium effort levels, and can match Opus 4.8 on some tasks at higher effort.
Availability
Sonnet 5 is live across all plans now. It's the default model for Free and Pro, and available to Max, Team, and Enterprise users. It's also in Claude Code and on the Claude Platform under the API name claude-sonnet-5.
Early testers consistently note that Sonnet 5 follows through on complex tasks better — where previous Sonnet models would stall halfway, Sonnet 5 checks its own output and pushes ahead without being told. Early access partners include teams from Glean, Lovable, Mercari, and Writer.
Claude Science: an AI workbench for researchers
Also out yesterday is Claude Science, an AI workbench for scientific researchers. It consolidates the fragmented tools scientists typically juggle — PubMed, Jupyter, R, cluster terminals, and more — into a single environment.
Research work involves a lot of tedium. Dozens of databases with different schemas, file formats that need custom pipelines, and constant context-switching between tools. Claude Science pulls these together so you can analyze literature, run multi-step studies, and produce publication-ready figures and papers from one workspace.
Every output carries a full audit trail: the code used, the environment it ran in, and what was changed. Results can be validated and reproduced.
Key features:
- 60+ pre-built skills covering genomics, single-cell analysis, proteomics, structural biology, cheminformatics
- Built-in Reviewer agent that checks citations and calculations
- Runs locally, via SSH to HPC clusters, or on cloud GPUs through Modal
- Natively renders 3D protein structures, genome browser tracks, and molecular structures
- Every figure ships with the code and environment that produced it
Claude Science launches in beta today for Pro, Max, Team, and Enterprise users.
Community reaction
On HN, Sonnet 5 hit 870 points and Claude Science 362 points — developer interest is strong. The pricing strategy has been a talking point too, with some joking that Anthropic has finally learned to compete on price.
The product lineup is becoming clearer: Opus handles frontier general capabilities, Sonnet tackles high-value agentic applications, and vertical products like Claude Science and Claude Code build moats in specific domains.