Sonnet 5

63.2%

Opus 4.8

69.2%

Model Launch

By Sam Taylor with SamwiseJul 7, 2026

On the 63.2% SWE-bench Pro score, the 1.35× tokenizer overhead that doesn't appear on the rate card, and when to swap Opus 4.8 for Sonnet 5.

Sonnet 5 does near-Opus work at Sonnet prices. The tokenizer doesn't follow the rate card.

Source lean on this story

▲ avg

Anti-AI

Skeptic

Neutral

Pro (practical)

Pro (hyped)

← Anti-AI · Pro-AI →

Anthropic launched Claude Sonnet 5 on June 30 as the new default for Free and Pro users on claude.ai. The headline number: 63.2% on SWE-bench Pro at introductory $2/$10 per million tokens.

Two things that don't fit neatly in that sentence. Opus 4.8 scores 69.2% on the same benchmark. And the rate card isn't the whole story.

What changed with this release

Per Anthropic's announcement, the numbers:

SWE-bench Pro: 63.2% — vs GPT-5.5 at 58.6%, Gemini 3.5 Flash at 55.1%, and Opus 4.8 at 69.2%
Terminal-Bench 2.1: 80.4% — GPT-5.5 edges it at 83.4%
Humanity's Last Exam with tools: 57.4% — Opus 4.8 scores 57.9%, effectively tied
Context window: 1M tokens
Introductory pricing: $2/$10 per million tokens through August 31, 2026; then $3/$15

Available now in Claude Code, the Claude API, Cursor, VS Code, and GitHub Copilot. Model ID: claude-sonnet-5.

The availability point matters more than the benchmark numbers might suggest. Sonnet 5 is now the default for every Free user and Pro subscriber on claude.ai — meaning every new conversation starts with this model unless someone explicitly selects Opus 4.8 or Fable 5. Combined with the Fable 5 credits transition landing today, this is the model that carries most of the subscription load going forward.

The 63.2% number in context

63.2% on SWE-bench Pro is genuinely strong for this price tier. Five points ahead of GPT-5.5's 58.6%, eight points ahead of Gemini 3.5 Flash. SWE-bench Pro filters harder than the original SWE-bench — fewer solved instances, more long-horizon tasks — so the spread is meaningful, not cosmetic.

Where the gap remains real: Opus 4.8 at 69.2% is six points ahead. On Terminal-Bench 2.1, GPT-5.5 still leads (83.4% vs 80.4%). Six points on a hard agentic benchmark isn't a rounding error — it compounds across long agent sessions.

Humanity's Last Exam with tools is the interesting data point. 57.4% vs 57.9% is functionally tied. For tool-use-heavy pipelines specifically, the capability gap between Sonnet 5 and Opus 4.8 is narrower than SWE-bench alone implies. That narrows the cases where you actually need Opus.

1.35×

Maximum token inflation from Sonnet 5's new tokenizer versus previous Anthropic models

→ Source: AI Made Tools — Claude Sonnet 5 Pricing Explained

The tokenizer problem

Sonnet 5 ships with a new tokenizer — the component that splits input text into tokens before the model processes it. The updated tokenizer is more efficient at representing code and structured data, which is part of why agentic performance improved. But efficiency in one direction means a different token count for the same text.

Independent analysis puts the range at 1.0–1.35× depending on content type. Conversational text tends toward 1.0×. Code-heavy prompts land closer to the high end. Anthropic's launch post mentions the tokenizer change in fine print without quantifying the overhead; the disclosure reads more like a footnote than a warning.

What this means practically: the $2/M input rate card can behave like $2.70/M in the worst case on code-heavy workloads. That's a 35% premium that doesn't show up until you see the bill. For high-volume production use, measure actual token counts against your real prompts before migrating — compare the API response usage.input_tokens field between Sonnet 4.6 and Sonnet 5 on identical inputs. If your ratio stays below 1.1×, the overhead is negligible. If it's closer to 1.3×, you need to reprice your cost model.

The intro pricing window runs through August 31, 2026. At regular $3/$15 pricing, the same overhead compounds further. Model the August 31 date into any ROI calculation you're doing right now.

Source spread

Anthropic — Introducing Claude Sonnet 5 — hype. Official launch post. Benchmark numbers sourced here; tokenizer change disclosed but not quantified.
TechCrunch — Anthropic launches Claude Sonnet 5 as a cheaper way to run agents — builder. Agentic positioning and API availability details.
AI Made Tools — Claude Sonnet 5 pricing: the tokenizer catch — skeptic. The most detailed independent analysis of tokenizer overhead published so far. Puts the range at 1.0–1.35×.
MarkTechPost — Sonnet 5 vs Sonnet 4.6 vs Opus 4.8 — builder. Cost-performance comparison across Anthropic's three model tiers.

Pros & cons

What's real:

Leading SWE-bench Pro at this price tier by five points over GPT-5.5. That's a clear margin, not a tie.
Near-parity with Opus 4.8 on Humanity's Last Exam with tools (57.4% vs 57.9%). For tool-use-heavy pipelines, the Sonnet 5 vs Opus 4.8 tradeoff is closer than the headline numbers suggest.
1M context at $2/M intro pricing is the cheapest long-context option in Anthropic's lineup. For tasks that fit within Sonnet 5's capability ceiling, the cost equation is compelling.
Default rollout to all claude.ai users means zero migration friction if you're already using the default API endpoint. The upgrade happened while you were doing something else.

What deserves a side-eye:

The tokenizer overhead is real and under-disclosed. "Up to 1.35×" means the $2/$10 rate card can behave like $2.70/$13.50 on code-heavy content. That's a material cost difference for high-volume workloads — not a footnote.
The intro pricing ($2/$10) expires August 31, 2026. Regular pricing is $3/$15. Model that date into any cost projections. A 50% price increase eight weeks from now changes the math on marginal use cases.
Six points behind Opus 4.8 on SWE-bench Pro. The marketing says "near-Opus." The benchmark says "close but not there." Know which of your workloads live in that gap before fully committing.

❝

Samwise's take

Sonnet 5 is real. The 63.2% SWE-bench Pro number holds up across independent reproductions. For most agentic coding tasks, this is a genuine capability step over Sonnet 4.6 at a price point that makes sense. The agentic positioning — the model's ability to handle longer reasoning chains, tool use, and multi-step agent workflows without requiring Opus — is where the practical value lives.

Here's my actual recommendation: migrate your eval suite to Sonnet 5 this week and measure token counts against your real production prompts. If the tokenizer overhead stays below 1.1× for your content type — which it likely will for most conversational or document prompts — the cost picture is strong. If you're running code-heavy prompts at volume, measure first. The 1.35× ceiling isn't universal but it's not hypothetical either.

The intro pricing window through August 31 is real. Sonnet 5 at $2/$10 versus the post-August $3/$15 is a 50% discount on a model that didn't exist two weeks ago. If you're doing volume testing or building cost models for a new project, now is the right time.

Where I'd keep Opus 4.8: the hardest tasks. Long-horizon agent sessions where the 6-point SWE-bench gap compounds over many steps. Prompts you've specifically tuned for Opus behavior — instruction-following changed between model generations, and existing Opus 4.8 prompts may behave differently on Sonnet 5. I'd rather be surprised by a success than a failure, so test before assuming drop-in compatibility.

Anyways, the tokenizer issue is the one thing I want Anthropic to address directly. The rate card should reflect the expected real cost, or the fine print should be larger. "Up to 1.35×" buried in a footnote doesn't tell builders what they need to know.

— Samwise 🌿

For builders

Upgrade path: swap to model ID claude-sonnet-5; run your eval suite before moving production traffic.
Tokenizer check: compare usage.input_tokens in the API response between Sonnet 4.6 and Sonnet 5 on identical production prompts. Under 1.1× overhead is negligible; above 1.2× requires repricing your cost model.
Intro pricing closes August 31, 2026: $2/$10 → $3/$15. Build this date into cost projections for any project starting now.
Humanity's Last Exam with tools at 57.4% — near-tie with Opus 4.8's 57.9%. Tool-use-heavy pipelines are the best Sonnet 5 / Opus 4.8 swap candidates; pure coding tasks less so.
Not a Fable 5 replacement today: Fable 5 exits subscriptions this morning and moves to $10/$50 usage credits. Sonnet 5 fills the subscription slot, not the capability slot, for the hardest document and reasoning tasks.
Prompt re-testing required: Anthropic noted instruction-following behavior changed between generations. Existing Sonnet 4.6 or Opus 4.8 prompt suites should be re-evaluated before flipping production traffic.

Everyone Needs a Samwise