Claude Sonnet 5 Drops at $2/$10 — But Read the Tokenizer Fine Print Before You Migrate

Hi guys, Mr. Technology here.

Anthropic shipped Claude Sonnet 5 on June 30, 2026, and it is, by the numbers, the most interesting mid-tier model release of the year. It is the default for Free and Pro accounts on day one. No waitlist, no preview, no partner gate. Just turn on Claude Code and it's there. That alone puts it ahead of GPT-5.6 Sol this week, which is sitting behind a government-approved partner wall of roughly 20 organizations.

So this is the real story: a near-Opus model that anyone can call today, at $2 per million input tokens and $10 per million output through August 31. After that it goes to $3/$15. Yes, those numbers look suspiciously like the old Sonnet 4.6 prices with a discount sticker slapped on top. We'll get to why that matters in a second. First, the good stuff.

The Headline Numbers

Sonnet 5 was built to be the most agentic Sonnet yet, and the benchmarks back that up. On SWE-Bench Pro — the harder of the SWE-bench variants, no public ground-truth leakage — Sonnet 5 lands at 63.2% pass rate. Sonnet 4.6 was 58.1%. Opus 4.8 is 69.2%. That's a 5-point jump over its own predecessor and a 6-point gap to the flagship. CursorBench mirrors it: 57% for Sonnet 5 versus 49% for Sonnet 4.6, the largest jump Cursor has reported between adjacent Sonnet releases.

Then there's Terminal-Bench 2.1, which is where things get weird in a good way. Sonnet 5 hits 80.4%. Opus 4.8 sits at 74.6%. The mid-tier beats the flagship on the same harness. Anthropic buried this in the system card rather than the launch blog, but it's right there in the data. If you're running command-line agents, this is the model you should be benchmarking first.

Anthropic's partner quotes are unusually consistent this cycle. Lovable said it "gets more done with less." ClickHouse said "reasons in tighter steps." Multiple teams described Sonnet 5 finishing end-to-end multi-step jobs that previous Sonnet models would stop halfway through. The pattern across the testimonials is the same: it doesn't need hand-holding to push through the boring middle of an agentic task.

What's Actually New Under the Hood

A few API-level changes that will hit you the day you migrate:

• 1 million token context window, 128K max output. Same ceiling as Opus, no more reason to pay the Opus tax just to fit a large repo in the prompt.
• Adaptive thinking is on by default. You can disable it with "thinking": {"type": "disabled"}, but expect to leave it on for anything agentic.
• temperature, top_p, and top_k are no longer supported. If your codebase is hard-coded around those sampling parameters, you have a migration to do. Anthropic is pushing everyone onto their effort-level dial instead.
• Cyber safeguards are enabled by default. This is why the US government didn't block the release. Sonnet 5 is significantly less capable at cyber tasks than Mythos 5, which is the model the export controls were designed around.

The Tokenizer Trap Nobody's Talking About

Here is the part that matters for your bill. Sonnet 5 ships with a new tokenizer, and the same input text produces roughly 30% more tokens than on Sonnet 4.6. Anthropic's own docs state it plainly.

Simon Willison ran the numbers against his Claude Token Counter. The deltas aren't uniform:

• English (Universal Declaration of Human Rights): 1.42x more tokens
• Spanish: 1.33x more tokens
• Python source (sqlite_utils/db.py): 1.28x more tokens
• Simplified Mandarin Chinese: 1.01x — basically flat

Do the math on that intro price. The headline says $2/$10 per million tokens. For an English-heavy workload, your effective price is closer to $2.84/$14.20 per million tokens. For a Python-heavy agentic workload, you're at $2.56/$12.80. After August 31, the post-intro price of $3/$15 becomes roughly $4.26/$21.30 effective on English.

This is not a price cut. It's a rebrand. Sonnet 4.6 was $3/$15 with a more compact tokenizer. Sonnet 5 at $3/$15 post-promo with a 30% larger tokenizer is essentially the same per-character cost, dressed up as a launch discount.

If you're running high-volume agentic workloads where the marginal token matters, benchmark before you migrate. The performance delta is real — Terminal-Bench 2.1 doesn't lie — but price your migration on measured token counts on your actual prompts, not the sticker price.

Who Can Use It and How

• Claude Free and Pro: Sonnet 5 is the default model. Just open the app.
• Max, Team, Enterprise: Available across all tiers in claude.ai and Claude Code.
• API: claude-sonnet-5 model ID on the Claude Platform, with the same tool and platform features as Sonnet 4.6.
• Pricing through August 31, 2026: $2 input / $10 output per million tokens.
• Pricing after August 31, 2026: $3 input / $15 output per million tokens.
• Context: 1M input, 128K output.

The Takeaway

Sonnet 5 is a legitimately better model. On Terminal-Bench 2.1 it crosses Opus 4.8. On SWE-Bench Pro it closes the gap from 11 points to 6. On CursorBench it posts the largest generation-over-generation jump Cursor has ever measured. If you're building agentic systems today and you weren't already routing to Claude, you should be testing Sonnet 5 this week.

But the pricing story is marketing, not math. The headline $2/$10 number assumes the old tokenizer. With the new tokenizer, your real per-character cost is roughly 30% higher than what the launch blog implies. Anthropic isn't hiding this — it's right there in the platform docs — but it's buried under enough launch hype that a lot of teams will miss it until the bill arrives.

Migrate, yes. But migrate with your token counter open.

— Mr. Technology

Model: Claude Sonnet 5 Lab: Anthropic Release date: June 30, 2026 Access: Default model for Claude Free and Pro; available on Max, Team, and Enterprise; available in Claude Code and on the Claude API as claude-sonnet-5 Pricing (intro, through Aug 31, 2026): $2 per million input tokens / $10 per million output tokens Pricing (after Aug 31, 2026): $3 per million input tokens / $15 per million output tokens Context window: 1,000,000 tokens input / 128,000 tokens output Effective price note: New tokenizer produces ~30% more tokens per text on English (~28% on Python), so effective per-character cost is roughly 1.3x the listed rate for English-heavy workloads.