Claude Sonnet 5 Review — Benchmarks, Pricing, and the Agent Angle
On June 30, 2026, Anthropic released Claude Sonnet 5 — the new mid-tier flagship in the Claude line, positioned less as “a smarter chatbot” and more as a cheaper way to run agents at scale. Unlike a preview drop, Sonnet 5 arrived with a full published benchmark card and confirmed pricing, so this review works from measured numbers rather than framing. The short version: it is a large step over the Sonnet 4.6 generation, it undercuts its own predecessor on introductory pricing, and its real pitch is the ratio of near-frontier reasoning to mid-tier cost.
TL;DR verdict
| Claude Sonnet 5 | |
|---|---|
| Type | Mid-tier frontier model (text + vision) |
| Released | June 30, 2026 |
| Context window | 1,000,000 tokens |
| Max output | 128,000 tokens (raisable to 300k via batch-API beta) |
| Pricing | Introductory $2 / $10 per 1M through Aug 31 → standard $3 / $15 |
| Headline numbers | SWE-bench Verified 72.7, GPQA Diamond 91.1 |
| New this release | New tokenizer (1.0–1.35× token count), default model on Free/Pro |
| Best for | High-volume agent pipelines that want near-frontier quality cheaply |
| Caveat | Below Opus 4.8 on the hardest coding; new tokenizer inflates token counts |
If you skip the rest: Sonnet 5 is the new default for cost-sensitive agent work. It closes much of the gap to the frontier on reasoning and agentic coding while staying at a mid-tier price, and Anthropic made it the default model on Free and Pro plans — a signal of where they expect most traffic to run. The two asterisks are the new tokenizer, which can quietly raise your token counts, and Opus 4.8, which still wins the hardest tasks.
What changed from Sonnet 4.6
The generational jump is real and concentrated on the axes that matter for agents:
- SWE-bench Verified: 72.7 — a large step up from the Sonnet 4.6 generation (low-60s) and the clearest signal that agentic, repository-level coding improved.
- GPQA Diamond: 91.1 — graduate-level science reasoning now sits close to the frontier tier, well above where the Sonnet line was.
- Terminal-bench: 76.1 — terminal-driven, tool-using coding jumped sharply from Sonnet 4.6’s mid-50s, which is exactly the workload agent frameworks lean on.
Alongside the raw scores, Sonnet 5 keeps the 1M-token context window and 128k max output (raisable to 300k tokens via a batch-API beta header), and it became the default model on Free and Pro plans at launch — Anthropic putting its highest-volume tier behind the model rather than reserving it for paying API customers.
What it costs
This is where the “run agents cheaply” framing gets concrete. Per 1M tokens:
| Window | Input | Output |
|---|---|---|
| Introductory (through Aug 31, 2026) | $2.00 | $10.00 |
| Standard (from Sep 1, 2026) | $3.00 | $15.00 |
The standard $3/$15 rate matches Sonnet 4.6, so once the promo ends you are paying the same list price for a materially stronger model — a straight capability upgrade at constant cost. During the introductory window it is cheaper than its own predecessor. Our models leaderboard records the durable standard $3/$15 rate rather than the promo, since the promo reverts in under two months; use the cost calculator to plug your own token mix against Opus 4.8 and the open coding tier.
One catch that changes the math: the new tokenizer. Anthropic notes the same input can map to 1.0–1.35× more tokens than on previous Claude models. A lower per-token price does not automatically mean a lower per-request bill — if your workload lands at the high end of that range, some of the introductory discount is absorbed by token inflation. Measure token counts on your own prompts before assuming the sticker price.
How it compares to Opus 4.8
Sonnet 5 does not dethrone Claude Opus 4.8 — and it is not meant to. Opus 4.8 stays ahead on the hardest agentic coding: it posts a higher SWE-bench Verified (88.6 on our leaderboard) and leads on SWE-bench Pro and OSWorld-Verified, where the extra compute shows up on long-horizon tasks. On the newer knowledge-work eval GDPval-AA, the two trade blows within a point.
The division of labor is clean:
- Opus 4.8 — the model for the hardest single tasks: deep refactors, long-horizon agent runs, anything where a quality cliff costs you more than the tokens.
- Sonnet 5 — the default for volume: most agent steps, tool calls, and everyday coding, at roughly a fifth of Opus 4.8’s output rate.
That mirrors the pattern the whole field has converged on — a frontier tier plus a cheaper workhorse under one family — and it is the same logic behind the tiering in OpenAI’s GPT-5.6 preview (Sol/Terra/Luna) and Google’s Flash/Pro split. Sonnet 5’s edge is that it shipped as a finished model with published numbers, not a preview.
Where it lands in the market
Sonnet 5 arrives in a crowded window. Google’s Gemini 3.5 Pro — announced at I/O in May with a 2M-token context and a “Deep Think” mode — remains in limited preview and has slipped to a July general-availability target, so as of this writing it is not yet a shipping competitor with published pricing. On the open side, coding-focused models like GLM-5.2 and Kimi K2.7-Code keep undercutting closed pricing on agentic coding, which is precisely the pressure a cheaper Sonnet answers.
Against that backdrop, Anthropic shipping a stronger-and-cheaper mid-tier — and making it the default on consumer plans — is a direct play for the high-volume agent market rather than the benchmark crown. The LLM Benchmark Comparison 2026 has the full cross-vendor picture if you want to weigh Sonnet 5 against the field on a per-benchmark basis.
Who should care
- Teams running agents at volume: This is the headline audience. If Sonnet 4.6 or an open coder was your default step-model, Sonnet 5 is a straightforward upgrade — better SWE-bench and Terminal-bench at the same standard price, cheaper during the promo. Validate the tokenizer’s token counts on your traffic first.
- Cost-sensitive pipelines currently on Opus: If you routed everything to Opus 4.8 for safety, Sonnet 5 is the tier to demote the easy steps to. See multi-agent pipelines for where a cheaper step-model slots into a larger workflow.
- Anthropic API users on Sonnet 4.6: Plan the swap, but re-test any prompt scaffolding that hard-codes token budgets or relies on 4.6’s verbosity — the new tokenizer changes both.
- Anyone ranking models today: Sonnet 5 has real published numbers, so it is rankable now — unlike the Gemini 3.5 Pro preview or the GPT-5.6 preview. It sits just below the closed frontier on capability and well below it on price.
FAQ
How much does Claude Sonnet 5 cost? Introductory $2 input / $10 output per 1M tokens through August 31, 2026, then standard $3 / $15 from September 1 — the same list rate as Sonnet 4.6. Our leaderboard records the durable $3/$15.
How does it compare to Opus 4.8? Opus 4.8 wins the hardest agentic coding (higher SWE-bench Verified and SWE-bench Pro); Sonnet 5 wins on cost, delivering near-frontier reasoning (GPQA Diamond 91.1) at roughly a fifth of Opus 4.8’s output price.
What is the new tokenizer? Sonnet 5 uses a new tokenizer that can map the same input to 1.0–1.35× more tokens than previous Claude models, affecting per-request cost and context usage. Measure on your own workload.
What’s the context window? 1,000,000 tokens, with 128k max output (raisable to 300k via a batch-API beta header).
Should I upgrade from Sonnet 4.6? For agentic coding and reasoning, yes — the gains are large and standard pricing is unchanged. Account for the tokenizer’s token inflation and re-test prompt scaffolding before switching production traffic.
Continue reading
- AI Models Leaderboard — Claude Sonnet 5 versus 60+ models on benchmarks, pricing, and context window, with a cost calculator.
- Claude Opus 4.8 Review — the frontier sibling Sonnet 5 sits just below, and the model for the hardest tasks.
- GLM-5.2 Review — the open coding flagship whose pricing is the pressure a cheaper Sonnet answers.
- GPT-5.6 Sol, Terra & Luna Review — OpenAI’s tiered family, the same flagship-plus-cheaper-tier pattern Sonnet 5 fits into.
- LLM Benchmark Comparison 2026 — the full cross-vendor benchmark picture.
- All Reviews — index of every head-to-head review on the site.