Sonnet 5's intro pricing ends August 31 — the config decides what you actually save

By Daniel S. · July 4, 2026

Anthropic launched Claude Sonnet 5 at an introductory price of $2 per million input tokens and $10 per million output — and that price has an expiration date: August 31, 2026. On September 1 it rises to $3/$15, a 50% jump on both numbers. That's the headline. The part that actually decides what you pay is the configuration around it, and one tokenizer footnote that most of the launch coverage skipped entirely.

Everything below is verified against Anthropic's pricing docs and launch announcement as of July 4, 2026. Prices move; if you're reading this later, check the source.

TL;DR

• Intro pricing: $2/$10 per MTok through August 31, 2026, then $3/$15. Batch API is half that: $1/$5 now, $1.50/$7.50 after. ⏱️ verified 2026-07-04
• Against Opus 4.8 ($5/$25), Sonnet 5 is 60% cheaper during the window, 40% cheaper after. Anthropic says performance is close to Opus on most work — close enough that routing matters more than model loyalty.
• The catch nobody quotes: Sonnet 5 uses a newer tokenizer that produces roughly 30% more tokens for the same text. Sonnet 4.6 costs the same $3/$15 Sonnet 5 will cost in September — but on the old tokenizer. Per-token prices aren't per-document prices.
• Three levers: set a default effort level and let hard tasks opt up, route to Opus on evidence rather than habit, and pull large one-time jobs into the window.
• Don't route security work here: Anthropic deliberately constrained Sonnet 5's cyber capability. That work stays on Opus.

The actual numbers

Source: Anthropic pricing docs, 2026-07-04. Sonnet 5 is the default model on Free and Pro plans and available on Max, Team, and Enterprise; the 1M-token context window is included at standard pricing.

The tokenizer footnote that changes the comparison

Buried in the pricing page: Sonnet 5 (along with Fable 5 and Opus 4.7+) uses a newer tokenizer that "produces approximately 30% more tokens for the same text." Anthropic credits it for part of the models' improved performance. Fine — but it means every per-token price comparison against older models is quietly off.

Run the math on the same document:

• Sonnet 5 at intro pricing ($2/$10, +30% tokens) works out to roughly $2.60/$13 in old-tokenizer terms — still cheaper than Sonnet 4.6's $3/$15, but the real discount on your actual text is closer to 13%, not 33%.
• Sonnet 5 at standard pricing ($3/$15, +30% tokens) is effectively more expensive per document than Sonnet 4.6 at the identical sticker price. You're paying for the capability gain, which may well be worth it — but know that's the trade you're making.
• Against Opus 4.8 the comparison is clean: Opus 4.8 uses the same new tokenizer, so 60%-cheaper-during-the-window means what it says.

The exact token increase depends on your content, so treat 30% as an estimate, not an invoice. But if your workload is high-volume text processing and Sonnet 4.6 already does the job, "upgrade to Sonnet 5" is not automatically the cheap option after August 31.

Lever 1: pull big jobs into the window

This is the only lever with a hard deadline. If you have large one-time work on the roadmap — a bulk classification run, generating descriptions for a catalog, a codebase-wide migration — running it before September 1 versus after is the same work at a 33% discount, purely on timing.

Stack the discounts and the gap widens. A batch job with 100M input and 20M output tokens:

• Before Aug 31, Batch API: 100 × $1 + 20 × $5 = $200
• After Aug 31, Batch API: 100 × $1.50 + 20 × $7.50 = $300

Same job, $100 apart. Prompt caching stacks on top of both (cache reads run at 0.1× the base input price), so a job with heavy repeated context drops further either side of the deadline — but the one-third timing gap stays.

Lever 2: an effort policy, not an effort habit

Sonnet 5 exposes adjustable effort — how hard the model thinks before answering. Anthropic's announcement keeps the details light ("select whichever level makes sense for your particular project"), but the cost mechanics are straightforward: more thinking means more output tokens, and output tokens are the expensive kind.

The mistake is picking one level and running everything at it. Cranked to maximum on every task, a cheap model stops being cheap — deep-reasoning runs can generate multiples of the output tokens a normal run produces, and at $10–15/MTok that adds up fast enough that a hard task at max effort can rival an Opus run at normal settings. Do that arithmetic on your own workload before assuming "high effort on Sonnet" beats "medium on Opus."

The fix is writing the policy down where the model reads it — in your CLAUDE.md — instead of deciding per task:

## Effort policy
- Default to medium effort.
- High effort only for: tricky debugging, multi-file
  refactors, architecture decisions.
- Low effort for: formatting, renames, boilerplate.

Now the cheap path is the default path, and the 20% of tasks that deserve deep thinking opt up explicitly.

Lever 3: route on evidence, not habit

Anthropic's own framing is that Sonnet 5 lands close to Opus 4.8 on most work. The launch charts show it matching Opus capability on some agentic tasks at a fraction of the cost — and the honest reading of "some" is that Opus still wins on the hardest reasoning.

So the cost-optimal setup isn't "switch everything to Sonnet 5." It's Sonnet 5 as default with a written escalation rule:

## Model routing
- Default: Sonnet 5 for coding, tool use, and day-to-day work.
- Escalate to Opus 4.8 when Sonnet has failed the same task
  twice, or the task needs the deepest reasoning.
- Security-sensitive work goes to Opus, always.

That last line isn't caution theater. Anthropic deliberately constrained Sonnet 5's cyber capability — their words: it "shows substantially poorer performance than models such as Opus 4.8" on those tasks, with safeguards on by default. If your work touches security analysis, that's a documented capability gap, not a routing preference.

Don't trust benchmarks for your own work, either. Take five real tasks you'd normally give Opus, run them on Sonnet 5 at medium effort, and compare. Whatever matches, route down permanently. Ten minutes of testing beats any launch chart, because it's measured on the work you actually do.

What I'd actually do before September 1

1. Set Sonnet 5 as the default model and write the effort policy into CLAUDE.md — ten minutes, one time.
2. Add the routing rule with the security carve-out.
3. Walk the roadmap for token-heavy one-time jobs and schedule them inside the window, on the Batch API.
4. If you're currently on Sonnet 4.6 with a high-volume text workload: run a real cost comparison on your own documents before migrating. The tokenizer change means the September price parity is not real parity.

FAQ

When does Sonnet 5 intro pricing end? August 31, 2026. From September 1 it moves from $2/$10 to $3/$15 per million input/output tokens — a 50% increase on both, confirmed in Anthropic's pricing docs.

Is Sonnet 5 cheaper than Sonnet 4.6? Per token during the intro window, yes. Per document, only modestly — Sonnet 5's newer tokenizer produces roughly 30% more tokens for the same text. At standard pricing the two models cost the same per token, which makes Sonnet 5 effectively more expensive per document.

Should I move everything from Opus 4.8 to Sonnet 5? Most work, probably; all work, no. Anthropic positions Sonnet 5 close to Opus on everyday coding and knowledge work at 40–60% less. Opus still wins on the hardest reasoning, and Sonnet 5's cyber capability is deliberately constrained — security work belongs on Opus.

Does the Batch API discount stack with the intro pricing? Yes. Batch runs at half price: $1/$5 per MTok until August 31, $1.50/$7.50 after. Prompt caching multipliers stack on top of both.

Sources: Introducing Claude Sonnet 5 (Anthropic); Claude Platform pricing docs. All figures verified 2026-07-04; pricing claims in this post expire with the intro window — if it's past August 31, 2026, read the table as history.