Claude Sonnet: the full family history

Sonnet is the Claude tier most people actually use. The complete release history from Claude 3 Sonnet to Sonnet 5, with specs, prices and retirement dates for each.

Sonnet is the middle tier of the Claude range and the one most people actually use. The current member is Claude Sonnet 5, released 30 June 2026, at $2 per million input tokens and $10 per million output tokens. Three Sonnet versions remain callable in August 2026 - Sonnet 5, Sonnet 4.6 and Sonnet 4.5 - while Sonnet 4 and Claude 3.7 Sonnet have been retired.

What follows is the full release history for the tier, version by version, with the specifications and prices Anthropic documents, an honest account of what changed at each step, and the reason Sonnet displaced Opus as the model of first resort for most work.

#Which Sonnet model is current, and what does it cost?

Sonnet 5, at $2/$10 per million tokens. That price launched as an introductory rate due to expire on 31 August 2026, and a rise to $3/$15 was scheduled for 1 September. On 10 August 2026 Anthropic cancelled the increase and made $2/$10 the standard price. Sonnet 5 is consequently the only current Claude model that costs less than the version it replaced.

One caveat spoils the headline slightly. Sonnet 5 uses the tokenizer introduced with Opus 4.7, which emits roughly 30% more tokens for the same text than Sonnet 4.6 did. A third off the sticker price against a third more tokens means real-world savings are real but smaller than they look, and the exact effect depends on your content. Re-measure rather than assume - the arithmetic is worked through on the Sonnet API page.

#Every Claude Sonnet version, in order

Where Anthropic's current documentation does not publish a figure for an older version, this table says so instead of reconstructing one from memory or from third-party write-ups.

^* Figures marked with an asterisk are the model's specifications at launch, recorded here as history: Claude 3.7 Sonnet and Sonnet 4 are retired, and Anthropic's current model and pricing pages no longer publish figures for them. The fuller records are on the Claude 3.7 Sonnet page and the Claude Sonnet 4 page.

#Why Sonnet became the default tier most people use

For the first eighteen months of the Claude 3 range, tier choice was simple: Opus was the capable one, Sonnet the compromise. Three things dissolved that.

#Sonnet caught up on the benchmark everyone cited

In September 2025, Sonnet 4.5 scored 77.2% on SWE-bench Verified. The then-current flagship, Opus 4.1, scored 74.5% - while costing $15/$75 against Sonnet's $3/$15. For the single most-quoted coding benchmark of that period, the mid-tier model was ahead of the premium one at a fifth of the output price. SWE-bench Verified is no longer reported by frontier vendors and should be treated as historical, but the commercial signal it sent in 2025 is why so many teams standardised on Sonnet and never moved back.

#The price kept going the right way

Sonnet held $3/$15 from Sonnet 4 through Sonnet 4.6, then fell to $2/$10. Batch processing halves that again to $1/$5, and cache reads cost a tenth of base input. At those rates the cost objection that pushes people down to Haiku 4.5 mostly evaporates for anything that is not genuinely high-volume.

#Opus is rationed on subscriptions, Sonnet is not

On Pro, Max, Team and Enterprise plans, Opus models are tracked on a separate weekly counter. Sonnet draws only on the shared allocation. In practice that means subscribers who lean on Opus run out of Opus first and finish the week on Sonnet - a structural nudge toward the middle tier that has nothing to do with capability. The pricing page explains how the two overlapping windows work; on the free plan the question is moot, because model selection is not offered at all.

#What changed between Sonnet generations

The tier's trajectory is easiest to read as three eras. Through Claude 3.5, progress was measured in raw task scores - the upgraded 3.5 Sonnet of October 2024 took SWE-bench Verified from 33.4% to 49.0%, and shipped alongside the first public beta of computer use. Claude 3.7 Sonnet in February 2025 reached 63.7%, or 70.3% with custom scaffolding, and arrived with the research preview of Claude Code; that version also became the subject of Anthropic's long-running Pokémon benchmark experiment.

The second era is agentic. Sonnet 4, launched with Opus 4 in May 2025 as part of Claude 4, scored 72.7%. Sonnet 4.5 pushed to 77.2% and 61.4% on OSWorld, still inside a 200k context window with extended thinking under a manual budget_tokens control.

The third era is architectural. Sonnet 4.6 brought the 1M-token window and 128k output to the tier, added adaptive thinking, and removed assistant message prefilling - a 400 error if you send it. Anthropic later restated Sonnet 4.6's OSWorld-Verified score as 78.5% and its Humanity's Last Exam results as 34.6% without tools and 46.8% with them. Sonnet 5 then turned thinking on by default, deleted the old manual extended-thinking mode outright, began rejecting temperature, top_p and top_k, and became the first Sonnet with real-time cybersecurity safeguards that can return stop_reason: "refusal" as a successful HTTP 200.

#When Sonnet is the wrong choice

Two directions. Upward: for multihour autonomous coding, large-scale refactoring or anything where a failure costs more than the tokens, the Opus tier and Fable 5 above it exist for a reason. Downward: for classification, extraction, routing and sub-agent work at volume, Haiku is half the price again and faster. The tier-by-tier comparison puts numbers on where the crossover falls.

There is also a documented gap worth knowing before you commit. On the GDM-MRCR v2 long-context retrieval evaluation, Gemini scores 97.0% against Sonnet 5's 81.5% - a million-token window is not the same thing as reliable recall across it. And Priority Tier is not available on Sonnet 5, which matters if you were relying on it for latency guarantees. The rest of the lineup is indexed on the Claude models page.

#Frequently asked questions