Anthropic ships Claude Sonnet 5, narrowing the gap to Opus

Anthropic shipped Claude Sonnet 5 on June 30, 2026, the company's first mid-tier model built for sustained agentic work. Sonnet 5 is the default on Free and Pro plans from launch day, sits close to Opus 4.8 on most capability evals, and starts at $2 per million input tokens for the first two months, before settling at $3 per million input and $15 per million output on September 1.

The launch lands the day after Anthropic published the Sonnet 5 system card and a separate post on its cyber safeguards rollout, and it is the first Sonnet release to ship with the same real-time cyber abuse detection that Anthropic ships in Opus 4.7 and 4.8. Anthropic shipped Sonnet 5 with cyber safeguards enabled by default because, in the company's own words, "we judged that the overall level of cybersecurity risk from Sonnet 5 was low." That is a notable downshift from the Fable 5 export-control rollout two weeks earlier, which Anthropic paused after the Commerce Department flagged the model for downstream cyber risk.

Sonnet 5 is also the first Sonnet that Anthropic explicitly positions against its own Opus line. The cost-performance curve in Anthropic's launch post places Sonnet 5 close to Opus 4.8 on agentic search and computer use benchmarks at medium effort, and the company framed the new effort level controls as a way to trade cost for capability within the same model. Opus 4.8 is the recommended choice for cybersecurity work that needs the relaxed guardrails.

The engineering and pricing changes inside Sonnet 5

The most concrete engineering change in Sonnet 5 is a new tokenizer. The same input maps to roughly 1.0 to 1.35 times as many tokens as it did under Sonnet 4.6, depending on content type. Anthropic is keeping the introductory pricing at $2 per million input tokens through August 31, 2026 so that the transition is roughly cost-neutral for the typical developer. The new standard price of $3 per million input tokens is below the $5 per million input that Opus 4.8 currently sells at.

Sonnet 5 is also the first Sonnet to ship with adjustable effort levels. Anthropic's own cost-performance charts show that the gap between Sonnet 5 at high effort and Opus 4.8 narrows on some agentic search and computer use tasks, while the gap between Sonnet 5 at low effort and Sonnet 4.6 is wide. The rate limits were also raised across Chat, Cowork, Claude Code, and the Claude Platform on April 26, 2026, ahead of this release, to accommodate the higher token usage of higher effort levels.

The capabilities story is more uneven. Anthropic published a single chart comparing Sonnet 5, Sonnet 4.6, and Opus 4.8 on BrowseComp, the agentic search benchmark, and on the OSWorld computer use benchmark, and then walked the BrowseComp chart back the same day in an edit note, citing a methodology mismatch that had understated Sonnet 5's performance. The corrected version uses a 10M token budget with compaction and programmatic tool calling, which is the same methodology Anthropic uses in the system card. The OSWorld-Verified score for Sonnet 4.6 was also updated to 78.5% under the revised methodology, up from the 34.6% number in the original Sonnet 4.6 launch post. The full evaluation set is in the system card.

Sonnet 5 is in the same Anthropic Cyber Verification Program as the Opus 4.x line, which means organizations that are already enrolled get Sonnet 5 access on the native Claude Platform, the Claude Platform on AWS, and Claude in Microsoft Foundry, with Vertex support coming. The guardrails are the same as Opus 4.7 and 4.8, not the stricter guardrails from Fable 5. The recommended use case for cybersecurity work that needs reduced guardrails is still Opus 4.8.

How Sonnet 5 lands against the 2026 model market

Sonnet 5 is the second major frontier release of June 2026, landing the same week that the broader AI agent infrastructure stack is consolidating. The pricing is the most disruptive piece for mid-market developers, who now have a model that runs sustained agentic workflows at close to Opus capability for roughly 40% of the per-token cost. The earlier Anthropic Claude Opus 4.7 release from April 2026 already moved the floor on enterprise coding tasks, and Sonnet 5 extends that floor down to the default-tier model.

The competitive read across the cheapest AI model APIs for startups in 2026 resource page is that Anthropic is now pricing against both the open-weight and the closed-weight competitors on a per-token basis at a level that the same page did not have to track a year ago. Whether the same pricing holds in the next six months depends on how the underlying inference economics shake out. The same week's Micron + Anthropic strategic agreement on HBM and Claude deployment gives the company a long-term memory supply line, which is the most direct lever on inference cost in the current generation of models.

The system card also reports that Sonnet 5 has a lower rate of misaligned behavior than Sonnet 4.6 in Anthropic's internal safety evaluations, and a much higher rate than Mythos Preview or Opus 4.8. Anthropic did not train Sonnet 5 on cybersecurity tasks, and on a Firefox exploit development evaluation the model scored 0% on full working exploits, the same as Sonnet 4.6, and a slightly higher partial success rate. The overall judgment in the system card is that the cyber capabilities are below Opus 4.8 and Mythos 5, and that the safeguards shipped in Sonnet 5 are appropriate for the current capability level. The full safety assessment is in the system card, which Anthropic published the day before the model itself went live.

The positioning for product teams is the most concrete part of the launch. For coding tasks where the model is going to be exercised as a long-running agent, the cost equation now reads as roughly the same per-run as Sonnet 4.6 because of the tokenizer adjustment and the intro pricing. For one-shot completions, the price is the same as the older Sonnet pricing on a per-token basis. The model that wins for a given workload is now Opus 4.8 if the workload is cyber-sensitive, Sonnet 5 if the workload is agentic but not cyber-sensitive, and Sonnet 4.6 if the workload is short and price-sensitive. The throughline is that the default tier keeps moving up the capability curve, and the user-visible surface keeps getting simpler.