Claude Sonnet 5

Overview

Claude Sonnet 5, released on June 30, 2026, is Anthropic's best combination of speed and intelligence. It substantially improves on Sonnet 4.6 for coding and agentic work, reaching what was previously Opus-tier quality on many tasks while staying in the Sonnet price band.

Anthropic calls it "the most agentic Sonnet model yet": it plans and uses tools autonomously at levels that previously required larger models. Its API surface has also converged with Opus 4.7/4.8 — manual extended thinking is gone, non-default sampling parameters are rejected, and adaptive thinking is on by default.

For most teams, Sonnet 5 is the workhorse: Opus 5 for the hardest problems, Haiku 4.5 for latency-critical paths, Sonnet 5 for everything in between.

Capabilities

• Agentic execution: Makes plans and uses tools autonomously; reaches for tools and self-verification loops more readily than Sonnet 4.6.
• Coding: The largest gains over Sonnet 4.6 are in coding and agentic tasks. Performs well out of the box on existing Sonnet 4.6 prompts.
• Agentic search: Anthropic's cost-performance curves say Sonnet 5's "higher-effort performance can match Opus 4.8 on some tasks" — a claim spanning the BrowseComp and OSWorld-Verified charts, not a BrowseComp-specific result.
• Computer use: Supports the computer_20251124 tool version. Cost-performance on OSWorld-Verified improves substantially over its predecessor.
• High-resolution vision: The first Sonnet-tier model to accept images up to 2576 pixels on the long edge.
• Progress reporting: Provides regular, higher-quality interim updates through long agentic traces by default — forced "summarize every N tool calls" scaffolding is now redundant.

Technical Specifications

• API model ID: claude-sonnet-5
• Context window: 1M tokens
• Max output: 128K tokens per request. On the Message Batches API, up to 300K via the output-300k-2026-03-24 beta header.
• Thinking: Adaptive thinking, on by default when the thinking parameter is omitted
• Effort levels: low, medium, high (default), xhigh, max
• Training data cutoff: January 2026 (reliable knowledge cutoff: January 2026)
• Tokenizer: The tokenizer introduced with Opus 4.7 — approximately 30% more tokens for the same text than Sonnet 4.6
• Comparative latency: Fast (faster than Opus 5 and Fable 5, slower than Haiku 4.5)
• Architecture: Transformer-based, developed under Anthropic's Constitutional AI approach

Supported features

Effort, adaptive thinking, compaction (beta), context editing (beta), the memory tool, code execution, programmatic tool calling, computer use (beta), high-resolution vision, prompt caching, Message Batches, and structured outputs.

Not supported: task budgets — Anthropic's docs list that beta on Fable 5, Mythos 5, Opus 4.8, and Opus 4.7, and mark Sonnet 5 "Not supported." Priority Tier is also unavailable on Sonnet 5.

API Behavior

Sonnet 5 brings the Sonnet line onto the same request surface as Opus 4.7/4.8. Anthropic names two breaking changes versus Sonnet 4.6:

• thinking: {"type": "enabled", "budget_tokens": N} — removed. The transitional escape hatch that still worked on Sonnet 4.6 is gone. Use thinking: {"type": "adaptive"} plus output_config.effort.
• Non-default temperature, top_p, top_k — rejected. Omitting them, or passing the default, is still accepted. Steer with prompting.

A third rejection is easy to misread as new: last-assistant-turn prefills return a 400 — but they already did on Sonnet 4.6. If you removed prefill during that migration, nothing further is needed. Use structured outputs or a system-prompt instruction.

Two silent default changes matter as much as the errors:

• Adaptive thinking is on when thinking is omitted. On Sonnet 4.6 the same request ran without thinking. Because max_tokens caps total output including thinking, a workload that relied on omission may now truncate.
• thinking.display defaults to "omitted" (it was "summarized" on Sonnet 4.6). Thinking blocks stream with empty text — to a streaming UI this reads as a long pause before output. Pass "summarized" explicitly if you surface reasoning.

One constraint that trips up tool-use code, and it is not platform-specific: with thinking active, tool_choice supports only {"type": "auto"} (the default) or {"type": "none"}. Forcing a tool with {"type": "any"} or {"type": "tool", "name": "..."} errors, because forced tool use is incompatible with extended thinking. Since adaptive thinking is now on by default, code that forced a tool on Sonnet 4.6 without thinking will fail on Sonnet 5 unless you pass thinking: {"type": "disabled"}.

Choosing an Effort Level

effort defaults to high. Sonnet 5 respects effort levels strictly, especially at the low end — at low and medium it scopes work to what was asked rather than going above and beyond.

If you observe shallow reasoning on complex problems, raise effort rather than prompting around it. At xhigh or max, leave generous max_tokens headroom — under the new tokenizer, limits tuned for Sonnet 4.6 may truncate equivalent output with stop_reason: "max_tokens".

Behavioral Notes

• More literal instruction following. Sonnet 5 does not silently generalize an instruction from one item to another. If a rule should apply broadly, state the scope explicitly. Holdover style directives from Sonnet 4.6 may now over-apply — re-baseline them.
• More agentic by default. It reaches for tools and self-verification more readily. With thinking disabled, however, it is less likely to reach for tools — add an explicit nudge if your harness depends on tool calls in that mode.
• Response length calibrates to task complexity rather than defaulting to a fixed verbosity.
• Code-review harnesses may show lower measured recall. Sonnet 5 follows "only report high-severity issues" more faithfully than earlier models. Tell it to report everything with confidence and severity, then filter downstream.

Use Cases

• Autonomous software development: Agentic coding at high throughput, with xhigh effort for the hardest tasks.
• Enterprise automation: Multi-step document processing, data extraction, and cross-system integration at scale.
• Computer use and browser automation: Real-world desktop and web tasks via the computer_20251124 tool.
• Agentic search and research: At higher effort, Anthropic says performance "can match Opus 4.8 on some tasks."
• Customer-facing assistants: Fast latency with near-frontier reasoning, at low/medium effort for cost control.
• Vision-heavy pipelines: Screenshots, charts, and document understanding at up to 2576px.

Performance / Benchmarks

Anthropic's launch announcement presents Sonnet 5 through cost-performance curves and a comparison table, both published as images. The only scores in the post's own text are restatements of Sonnet 4.6 figures. The numbers below are transcribed from Table 8.1.A, "Capability evaluation summary," in the Claude Sonnet 5 System Card (June 30, 2026), which publishes them as text.

The caption gives the setting for every row unless otherwise noted: "adaptive thinking at max effort, default sampling settings (temperature, top_p), averaged over 5 trials." Context is evaluation-dependent — "Standard configurations use 1M tokens; BrowseComp uses a 10M-token limit with context compaction (triggered at 200k)." These are max-effort scores, not the high default.

The Terminal-Bench harness changed, and it matters. Sonnet 5's 80.4 was measured with mini-SWE-agent, not the Terminus-2 harness behind Opus 4.8's system-card figure. Anthropic's reason: at xhigh effort, "Terminus-2 experiences 2.7× more timeouts than mini-SWE-agent ... this makes final scores noisier and less legible." Sonnet 5's 80.4 and Opus 4.8's system-card 74.6 therefore do not show Sonnet 5 outscoring Opus 4.8 — they are different harnesses.

Two footnoted restatements in the launch post revise Sonnet 4.6, not Sonnet 5. Anthropic "updated the grader model for Humanity's Last Exam" and restated Sonnet 4.6 at 34.6% (no tools) and 46.8% (with tools); it also changed how it runs OSWorld-Verified and updated Sonnet 4.6 to 78.5%.

Cost-performance curves

The announcement's headline framing is comparative rather than numeric. On BrowseComp (agentic search) and OSWorld-Verified (computer use), Sonnet 5 is "a strict improvement over Sonnet 4.6" and "covers a much wider range of cost-performance options than Opus 4.8"; "its higher-effort performance can match Opus 4.8 on some tasks." Those curves are images, and no per-point values are published as text.

Artificial Analysis

Sonnet 5's Artificial Analysis page reads: "Claude Sonnet 5 (Adaptive Reasoning, Max Effort) scores 53 on the Artificial Analysis Intelligence Index." The page labels that index v4.1; scores are not comparable across index versions.

How to read these numbers

All of the above are vendor-reported. Anthropic chose the benchmarks, the harness, and the effort level — max effort throughout, not the high default that production traffic actually runs at. Treat them as an upper bound and validate on your own evaluation set.

On safety, Anthropic reports a nuanced result, and the direction matters: on its automated behavioral audit Sonnet 5 "scored lower (that is, safer) overall" than Sonnet 4.6, but "did show somewhat higher rates of misaligned behavior on this assessment compared to the more capable Opus 4.8." Separately, it "has a much lower ability to perform cybersecurity tasks than our current Opus models." No ASL designation appears in the announcement; the system card is dated June 30, 2026.

Limitations

• Token counts rose ~30% for the same text versus Sonnet 4.6, under the new tokenizer. Per-token pricing is unchanged, so an equivalent request can cost more. Re-baseline with count_tokens before reacting.
• Reduced parameter surface: no budget_tokens, no non-default sampling parameters, no assistant prefill.
• Less tool-eager with thinking disabled — an explicit prompt nudge is needed if your harness relies on tool calls in that configuration.
• Cybersecurity refusals: Sonnet 5 is the first Sonnet-tier model with real-time cybersecurity safeguards, so requests touching prohibited or high-risk topics may be declined where Sonnet 4.6 answered. This is a safeguards change, not a capability one — Anthropic reports Sonnet 5 has "a much lower ability to perform cybersecurity tasks" than the current Opus models.
• Image token cost: High-resolution images can consume up to roughly 3× more image tokens than on Sonnet 4.6.
• No internet access: Cannot browse unless integrated through tool use or the server-side web search tool.

Pricing & Access

API Pricing

Available on the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud, and Microsoft Foundry:

• Input: $3 / MTok — $2 / MTok introductory through August 31, 2026
• Output: $15 / MTok — $10 / MTok introductory through August 31, 2026
• Prompt caching: cache reads at roughly 10% of base input price
• Batch API: 50% discount

Consumer Plans

Available on Free, Pro, Max, Team, and Enterprise plans, and in Claude Code. Anthropic states it is the default model for Free and Pro plans.

Ecosystem & Tools

Official SDKs

• Python SDK — full-featured with async support
• TypeScript SDK — type-safe for Node.js and browsers
• REST API — direct HTTP access from any language

SDKs also ship for Java, Go, Ruby, C#, and PHP, along with the ant command-line interface.

Developer Tools

• Anthropic Console — playground and prompt development
• Claude Code — agentic coding in the terminal, IDE, and web
• Claude.ai — browser-based chat

Community & Resources

• Introducing Claude Sonnet 5 — official announcement
• Claude Sonnet 5 System Card — safety and capability evaluation, June 30, 2026
• Models Overview
• Migration Guide — Sonnet 4.6 → Sonnet 5 breaking changes
• Pricing Page