Modelgeneral LLM4w ago

Claude Sonnet 5 review

Anthropic's Sonnet-tier model — 'the best combination of speed and intelligence.' Near-frontier on coding and agentic work at roughly a third of Opus's price, with a 1M-token context and adaptive thinking on by default.

By Ravi Menon · Models & Benchmarks EditorVerified 2026-07-24
Maker
Anthropic
Launched
Jun 30, 2026
Pricing
paid
Visit official site
Firstlook

Our verdict

Sonnet 5 is Anthropic's value pick: it reaches near-frontier quality on coding and agentic work while staying fast and cheap relative to Opus. Adaptive thinking is on by default, the context window is a full 1M tokens, and introductory pricing of $2 / $10 per MTok (through Aug 31, 2026) undercuts Opus 4.8 by roughly 2.5x. The catch is the same as every Claude model — it's closed and hosted-only, with no weights to run yourself — plus a newer tokenizer that counts ~30% more tokens than Sonnet 4.6. For most production workloads it's the sensible default; we haven't run it hands-on, so no score yet.

First look — our read from the docs and sources below; not yet hands-on tested.

Claude Sonnet 5 is Anthropic's middle tier — and, for a lot of teams, the one that matters most. Anthropic bills it as "the best combination of speed and intelligence": the model that gets close to frontier quality on coding and agentic work without Opus's price tag or latency. It sits below Opus 4.8 (the capability ceiling of the general lineup) and above Haiku 4.5 (the fastest, cheapest tier).

The headline spec sheet matches the current generation: a 1-million-token context window, up to 128K output tokens, a January 2026 knowledge cutoff, and adaptive thinking that's on by default — the model decides how hard to think, tuned by an effort dial that defaults to high. One migration wrinkle worth flagging: Sonnet 5 uses the newer tokenizer, which counts roughly 30% more tokens than Sonnet 4.6 for the same text, so any budgets carried over from an older Sonnet need re-baselining.

Who it's for

Most production workloads. If you're shipping a coding assistant, an agent, or a high-volume pipeline, Sonnet 5 is the default: near-frontier capability at a fraction of Opus's cost. Introductory pricing of $2 / $10 per million tokens (through August 31, 2026) makes it roughly a third the price of Opus 4.8, and it's faster by default, which shows up directly in interactive latency.

Who should skip it

Two cases. First, the hardest long-horizon agentic and enterprise tasks, where Opus 4.8's extra capability is worth the premium — reach up when the task genuinely needs it. Second, anyone who needs to own the model: Sonnet 5 is closed and hosted-only, so if on-prem, offline, or auditable weights are a hard requirement, an open-weight model is the only path.

This is a first look assembled from Anthropic's official documentation, not our own evaluation, so we're not attaching a score. On paper, Sonnet 5 is the value sweet spot of the Claude lineup — the model to start with, and to keep unless a task specifically demands Opus.

Provider

Provideranthropicclaude-sonnet-5· Proprietary (closed, hosted only)

Specs & key facts

What it isAnthropic's Sonnet-tier model — the best combination of speed and intelligence[src]
Context window1M tokens in · 128K out[src]
Pricing$2 / $10 per MTok intro (→ $3 / $15 from Sep 1 2026)[src]
ThinkingAdaptive thinking on by default (effort defaults to high); manual extended thinking not supported[src]
Knowledge cutoffJan 2026[src]
LicenseProprietary — hosted only, no downloadable weights[src]

Capabilities

CodingYes (near-frontier coding at Sonnet cost)
Reasoning / mathYes (adaptive thinking on by default)
Agentic / tool useYes
Self-hostNo (hosted only)
Hosted APIYes (Claude API)
Commercial useYes (commercial API terms)
Fast modeNo (fast mode is Opus 4.8 / 4.7 only)

How to use it

  1. 1Call it via the Claude API with the model ID `claude-sonnet-5` — hosted only, no weights to download.
  2. 2Adaptive thinking is on by default; set `effort` explicitly (it defaults to `high`) and reach for `xhigh` on the hardest coding/agentic tasks.
  3. 3Note the newer tokenizer produces roughly 30% more tokens than Sonnet 4.6 for the same text — re-baseline any token budgets.
  4. 4Use it as the default for most production workloads and step up to Opus 4.8 only when a task genuinely needs the extra capability.

Pricing

Claude API — introductory (through Aug 31, 2026)

$2 / $10 per MTok

$2 per million input tokens, $10 per million output tokens while introductory pricing is in effect. Cache reads bill at $0.20 / MTok.

Claude API — standard (from Sep 1, 2026)

$3 / $15 per MTok

Standard pricing takes effect September 1, 2026: $3 per million input tokens, $15 per million output tokens.

Closed, hosted model priced per token. Introductory pricing of $2 in / $10 out per million tokens runs through August 31, 2026, after which standard $3 / $15 pricing applies. The full 1M-token context window is billed at standard rates. Verified against Anthropic's docs 2026-07-24.

Pros & cons

Pros

  • Near-frontier coding and agentic quality at Sonnet-tier price and speed.
  • Introductory pricing $2 / $10 per MTok — about a third of Opus 4.8.
  • 1M-token context at standard rates, with adaptive thinking on by default.
  • Faster latency than Opus, making it a strong default for production.

Cons

  • Closed and hosted only — no downloadable weights to run or audit.
  • Newer tokenizer counts ~30% more tokens than Sonnet 4.6 for the same text.
  • No fast mode (that tier is Opus 4.8 / 4.7 only).
  • Introductory pricing ends Aug 31, 2026, after which it rises to $3 / $15 per MTok.

Alternatives

FAQ

Sources

  1. 1.Model ID, 'best combination of speed and intelligence' positioning, 1M context, 128K max output, adaptive thinking on by default, Jan 2026 knowledge/training cutoff, faster latencyhttps://platform.claude.com/docs/en/about-claude/models/overviewVerified 2026-07-24
  2. 2.Introductory pricing $2 / $10 per MTok through Aug 31 2026, standard $3 / $15 from Sep 1 2026; cache read $0.20 / MTok (intro)https://platform.claude.com/docs/en/about-claude/pricingVerified 2026-07-24
  3. 3.Release date June 30, 2026 (announcement + System Card)https://www.anthropic.com/news/claude-sonnet-5Verified 2026-07-24
  4. 4.Benchmarks (Table 8.1.A / §8.2): SWE-bench Verified 85.2%, SWE-bench Pro 63.2%, SWE-bench Multilingual 78.3%, Terminal-Bench 2.1 80.4%, OSWorld-Verified 81.2%, BrowseComp 84.7% single-agent, HLE 43.2% no tools / 57.4% with tools, AutomationBench 13.5%https://www.anthropic.com/claude-sonnet-5-system-cardVerified 2026-07-24

More coverage

News & first-looks about this release. Coming soon.
Head-to-head comparisons. Coming soon.