All models

Claude Sonnet 5

Anthropic's most agentic Sonnet, released June 30, 2026 — near-Opus quality on coding and agentic work at Sonnet-tier cost, with adaptive thinking on by default and the first Sonnet to support xhigh effort.

1M tokens · Text / Vision / Code · Prompt cache

Claude Sonnet 5 is Anthropic's most agentic Sonnet model, reaching what was previously Opus-tier quality on coding and agentic work at Sonnet-tier cost. It can plan, drive tools like browsers and terminals, and run autonomously at a level that until recently only larger Opus-class models could reach. It sits in the middle of the Claude 5 lineup — above Haiku 4.5, below Opus 4.8, and well below the premium Fable 5.

The API surface now aligns with the Opus tier: adaptive thinking is on by default (omitting thinking runs adaptive), manual thinking budgets are removed, and it is the first Sonnet to support the xhigh effort level. It uses a new tokenizer that produces roughly 1.0–1.35× more tokens for the same text than Sonnet 4.6, so re-baseline any token budgets when you migrate. Introductory pricing is $2/$10 per 1M tokens through August 31, 2026 (then $3/$15) — on VM0 it carries a ×1 baseline credit multiplier, the platform's balanced workhorse.

What is Claude Sonnet 5?

Released June 30, 2026 · The mid-tier of the Claude 5 family — the balanced default between Haiku 4.5 and Opus 4.8, and the cost-efficient workhorse beneath Fable 5. Ships at a ×1 VM0 multiplier.

Claude Sonnet 5 was released on June 30, 2026 as Anthropic's most agentic Sonnet to date, 21 days after Fable 5. Its largest gains over Sonnet 4.6 are in coding and agentic work — it reaches near-Opus quality while staying at Sonnet-tier cost, which makes it the natural default for most agent loops.

The API surface aligns with the Opus tier. Adaptive thinking is on by default (omitting the thinking field runs adaptive, where Sonnet 4.6 ran thinking-off), manual budget_tokens is removed, and non-default sampling parameters are rejected. Sonnet 5 is the first Sonnet-tier model to support the xhigh effort level, alongside low, medium, high, and max.

Sonnet 5 uses a new tokenizer that produces roughly 1.0–1.35× more tokens for the same text than Sonnet 4.6, so the 1M context window holds less text and token-budgeted limits shift even though the per-token rate is unchanged. It is also the first Sonnet-tier model with high-resolution vision, accepting images up to 2576 pixels on the long edge.

What's notable about Claude Sonnet 5

Headline architecture and capability features.

Sonnet 5 keeps the 1M-token context window and up to 128K output tokens. Adaptive thinking is on by default and controlled through effort — low, medium, high, xhigh, and max — with xhigh recommended for the hardest coding and agentic tasks. A new tokenizer produces ~1.0–1.35× more tokens per text than Sonnet 4.6. High-resolution vision accepts images up to 2576px on the long edge. Assistant-message prefill is not supported; use structured outputs instead.

Specs at a glance

FamilyClaude 5 generation
ModalitiesText, vision, code
LanguagesEnglish-first, multilingual
Prompt cachingSupported (Anthropic)
Context window1M tokens
Max outputUp to 128K tokens
Effort levelsLow / Medium / High / xhigh / Max
Vendor list price$2/$10 intro per 1M through Aug 31, 2026 (then $3/$15)

Claude Sonnet 5 benchmarks

Vendor-reported scores. Third-party trackers report different figures for the same benchmarks depending on dataset, context settings, and agent scaffold, so treat the headline numbers as configuration-dependent. Anthropic highlights the harder SWE-bench Pro set and the Terminal-bench jump as the signals that matter for agent builders.

SWE-bench Verifiedvendor-reported; up sharply from Sonnet 4.6
85.2%
SWE-bench Proclose to Opus 4.8's 69.2% at less than half the price
63.2%
Terminal-benchup from Sonnet 4.6's 55.4% — the headline jump for agent builders
76.1%
SWE-bench Multilingualvendor-reported
78.3%
SWE-bench Multimodalvendor-reported
28.1%

Claude Sonnet 5 pricing

Provider list price, per 1M tokens.

Input$2.00
Output$10.00
Cache read$0.20
Cache write$2.50

How Claude Sonnet 5 behaves in practice

Observed behaviour from production agent runs.

Near-Opus coding and agentic quality

Sonnet 5's biggest gains over Sonnet 4.6 are in coding and agentic tasks, reaching a level that until recently required Opus-class models. Its SWE-bench Pro score sits close to Opus 4.8's at less than half the cost — the reason it is the new default for most agent loops.

Adaptive thinking on by default

Omitting the thinking field now runs adaptive thinking, where Sonnet 4.6 ran thinking-off. Sonnet 5 calibrates thinking spend to task complexity and is the first Sonnet with the xhigh effort level — the recommended setting for the hardest coding and agentic work.

More agentic tool use

Sonnet 5 reaches for tools and runs self-verification loops more readily than Sonnet 4.6 by default, and provides higher-quality progress updates over long agentic traces. Raise effort to high or xhigh when you want even more tool usage in agentic search and coding.

New tokenizer — re-baseline budgets

The same text produces roughly 1.0–1.35× more tokens than on Sonnet 4.6, so the 1M window holds less and token-budgeted limits shift even though the per-token rate is unchanged. Re-measure representative prompts before reacting to any cost change.

Best agent tasks for Claude Sonnet 5

The everyday agent loop that used to need Opus

Most agent loops — reading files, editing code, driving tools, verifying output — run at near-Opus quality on Sonnet 5 for a fraction of the cost. Make it the default worker and promote only the hardest planning steps to Opus 4.8 or Fable 5.

The terminal-driven coding agent

The Terminal-bench jump to ~76% (from Sonnet 4.6's ~55%) is the headline for agent builders. Use Sonnet 5 with xhigh effort for autonomous, terminal-heavy coding where wall-clock cost matters.

The high-volume workhorse

At $2/$10 introductory pricing and a ×1 VM0 multiplier, Sonnet 5 is the cost-efficient default for the bulk of agent traffic — cheap enough to run everywhere while still handling hard reasoning and long-context recall.

When to skip Claude Sonnet 5

Skip Sonnet 5 on the very hardest reasoning or the longest autonomous runs where Opus 4.8 or Fable 5 pull clearly ahead, and on latency-critical chat where a cheaper model like Kimi K2.7 Code is faster. Note the new tokenizer produces more tokens for the same text than Sonnet 4.6, so re-baseline any tight token budgets before assuming cost parity.

Claude Sonnet 5 vs other models

Claude Sonnet 5 vs Claude Sonnet 4.6

Sonnet 5 is a substantial upgrade on coding and agentic work — near-Opus quality where 4.6 was a step behind, with a large Terminal-bench jump. The trade-offs: adaptive thinking is now on by default (4.6 ran thinking-off), and the new tokenizer produces ~1.0–1.35× more tokens for the same text, so re-baseline token budgets. Migrate new agents to Sonnet 5; keep 4.6 pinned only where you've validated against it.

Claude Sonnet 5 vs Claude Opus 4.8

Sonnet 5 (×1) trails Opus 4.8 (×2) on the hardest reasoning and longest-horizon runs, but closes most of the coding gap — its SWE-bench Pro sits close to Opus 4.8's at under half the price. The standard pattern is Sonnet 5 as the workhorse with Opus 4.8 as the planner for the hardest steps.

Claude Sonnet 5 vs Claude Fable 5

Fable 5 is the premium ceiling for the hardest reasoning and long-horizon autonomy; Sonnet 5 is the cost-efficient default. Use Fable 5 as an orchestrator that delegates fan-out work to cheaper Sonnet 5 sub-agents, and reserve its ×4 multiplier for where it actually changes the outcome.

Claude Sonnet 5 vs GPT-5.4 Mini

Both are balanced mid-tier picks. Sonnet 5 leads on agentic coding and long-context recall; GPT-5.4 Mini competes on raw token price. On VM0 the choice is a per-workload routing decision rather than a lock-in.

Bottom line: should you use Claude Sonnet 5?

The new default for most agents in the Claude 5 family. Near-Opus coding and agentic quality at Sonnet cost makes it the workhorse you run everywhere — promote to Opus 4.8 or Fable 5 only for the hardest reasoning, and keep a cheaper tier for latency-critical chat.

Frequently asked questions

When was Claude Sonnet 5 released?

Anthropic released Sonnet 5 on June 30, 2026, 21 days after Fable 5. It is available on VM0 today (model id claude-sonnet-5).

How much does Sonnet 5 cost?

Introductory pricing is $2 per 1M input tokens and $10 per 1M output through August 31, 2026, then $3/$15. On VM0 it carries a ×1 baseline credit multiplier — the balanced workhorse tier.

How is Sonnet 5's thinking different from Sonnet 4.6?

Adaptive thinking is on by default — omitting the thinking field now runs adaptive, where Sonnet 4.6 ran thinking-off. Manual budget_tokens is removed; control depth with the effort parameter. Sonnet 5 is also the first Sonnet with the xhigh effort level.

Why did my token counts go up after switching to Sonnet 5?

Sonnet 5 uses a new tokenizer that produces roughly 1.0–1.35× more tokens for the same text than Sonnet 4.6. The per-token rate is unchanged, but your actual per-task cost and context usage can be higher — re-measure representative prompts before reacting.

Does Sonnet 5 support high-resolution images?

Yes. Sonnet 5 is the first Sonnet-tier model with high-resolution vision, accepting images up to 2576 pixels on the long edge — useful for screenshots, charts, and document understanding.

Alternatives

Using Claude Sonnet 5 on VM0

Two ways to access Claude Sonnet 5 on VM0

VM0 supports Claude Sonnet 5 as a Built-in model billed in VM0 credits, and through bring-your-own with a Anthropic API key. The Built-in path uses VM0 Managed routing and the credit multiplier explained below; the bring-your-own path bills you directly with the upstream vendor and skips the VM0 credit conversion entirely.

VM0's recommendation

VM0 positions Claude Sonnet 5 as a core agent model, recommended alongside Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6 for the steps that drive the actual outcome of an agent run. These are the models we'd pick for the orchestrator role, for code-touching agents, and for any step where a wrong answer is expensive.

Credits and the ×1 multiplier

Every Built-in model on VM0 is priced as a multiple of Claude Sonnet 4.6, which sits at the ×1 credit baseline. Claude Sonnet 5 bills at ×1 credits. The multiplier is what shows up on your VM0 invoice; the vendor list price in the pricing table above is what the upstream provider charges before VM0 converts it into credits.

Claude Sonnet 5 sits at the ×1 baseline that every other Built-in model is priced against, so it's the unit you compare costs in when picking between models on VM0.

Available on VM0 since June 30, 2026.