Claude Sonnet 5 — Benchmark the Tokenizer Inflation Before Sept 1

Anthropic shipped Claude Sonnet 5 on June 30 with a new tokenizer that maps the same input to up to 1.35× more tokens than Sonnet 4.6 — and set introductory pricing at $2/$10 per million tokens specifically to keep your bill flat while it happens. Standard pricing ($3/$15) kicks in September 1. If you are running Claude Code at scale on Bedrock or Google Cloud, you have eight weeks to benchmark your real traces before the bill doubles.

TL;DR

• Tokenizer trap: Same input → up to 1.35× more tokens on Sonnet 5 vs Sonnet 4.6
• Introductory pricing ($2/$10) expires August 31 — standard $3/$15 starts September 1
• Benchmark gains are real: Terminal-Bench 2.1 jumps 13.4 points to 80.4%, FrontierCode more than doubles
• Action: Benchmark your actual traces NOW — follow the checklist below before the discount window closes

Claude Sonnet 5 — What Happened

Anthropic released Sonnet 5 as the default model for all Free and Pro Claude users, with availability across Max, Team, and Enterprise tiers. The headline numbers are strong: SWE-bench Pro moves from 58.1% to 63.2%, Terminal-Bench 2.1 leaps from 67.0% to 80.4% (the largest single-benchmark generational jump Anthropic has published), and FrontierCode v1 more than doubles from 15.1% to 38.8%. OSWorld-Verified nudges from 78.5% to 81.2%.

Buried further down in the announcement sits the tokenizer change. Sonnet 5 uses an updated tokenizer that improves performance but changes how text is processed. The same input that consumed 1,000 tokens on Sonnet 4.6 can consume up to 1,350 tokens on Sonnet 5. Anthropic’s response: introductory pricing at $2/$10 per million tokens (vs standard $3/$15) through August 31, explicitly designed to make migration “roughly cost-neutral.”

Simultaneously, Anthropic launched the Claude Apps Gateway — a self-hosted control plane for Claude Code on Bedrock and Google Cloud. It is a single stateless container backed by PostgreSQL, providing corporate SSO via OIDC, role-based model policies, per-user spend caps (daily/weekly/monthly), and usage telemetry export to a collector you operate. Inference traffic and usage data stay inside your cloud perimeter. Onboarding a developer means adding them to your IdP; offboarding means removing them.

Why This Matters

The math is not complicated, but it is worth stating plainly because Anthropic’s framing invites you to skip it.

During the introductory window, a request that previously cost you $3 on Sonnet 4.6 (at $3/MTok input) will cost roughly $2 × 1.35 = $2.70 on Sonnet 5 — cheaper, as promised. After September 1, that same request costs $3 × 1.35 = $4.05. That is a 35% effective price increase over Sonnet 4.6 for token-heavy workloads, wearing the disguise of “same list price.” For output tokens, the swing is even sharper: $15 × 1.35 = $20.25 effective per million output tokens vs Sonnet 4.6’s $15.

The 1.35× multiplier is a worst case — Anthropic states the range is 1.0–1.35× depending on content type. Code-heavy inputs with lots of syntax tokens tend to inflate more than natural-language prose. The teams hit hardest are exactly the ones running Claude Code in CI loops, agent chains, and automated refactoring pipelines — the teams Anthropic is most aggressively courting with the Apps Gateway launch.

I have watched this exact pattern play out with Opus 4.7 and Opus 4.8. Each time, the introductory discount kept bills flat long enough for teams to commit their tooling — CI integrations, agent orchestration configs, prompt libraries — and by the time standard pricing arrived, switching back was more expensive than absorbing the increase. Sonnet 5 runs the same play, but the blast radius is larger because it is now the default model for every Free and Pro user, not an opt-in upgrade.

The Apps Gateway’s spend caps meter at list price — committed-use discounts and negotiated rates may not be reflected in the gateway’s spend ledger. Use spend caps as a runaway-usage guardrail, not as bill reconciliation. You still need to reconcile actual invoices separately.

The xhigh Effort Cost Inversion

There is a second cost trap hiding in the effort levels. At xhigh effort, Sonnet 5 burns tokens on extended thinking to close the capability gap with Opus 4.8. On the hardest agentic tasks, it still does not close it — SWE-bench Pro lands at 63.2% vs Opus 4.8’s 69.2%, OSWorld at 81.2% vs 83.4%. But the token consumption at xhigh can push per-task cost above what Opus 4.8 charges at medium effort for comparable accuracy. Any team defaulting to xhigh for CI-style agent loops should run a direct cost comparison against Opus 4.8 at medium before committing.

Where Sonnet 5 Actually Wins

The benchmarks where Sonnet 5 beats Opus 4.8 outright are Terminal-Bench 2.1 (80.4% vs 74.6%) and GDPval-AA v2 knowledge-work (1,618 Elo vs 1,615). Terminal-Bench is the one that matters for Claude Code users: it tests CLI-based agent workflows — repo navigation, file manipulation, shell scripting, multi-step terminal operations. If your primary use case is a CLI agent that lives in the terminal rather than an IDE, Sonnet 5 is genuinely the better model, not just the cheaper one during the promo window.

FrontierCode’s jump from 15.1% to 38.8% is dramatic but the benchmark itself is new enough that the community has not stress-tested its correlation to real-world productivity. Treat it as a signal, not a decision driver.

Your Token Inflation Benchmark Checklist

1. Export 50–100 representative Claude Code requests from the past 30 days (mix of code-heavy and prose-heavy)
2. Replay each request through both Sonnet 4.6 and Sonnet 5 via the API — log usage.input_tokens and usage.output_tokens from each response
3. Compute your multiplier — divide Sonnet 5 token counts by Sonnet 4.6 counts per request to find your actual inflation factor (expect 1.0–1.35×)
4. Price it twice — multiply your Sonnet 5 token counts by $2/$10 (current) and $3/$15 (September) to see the real delta
5. Update routing policy — if September cost exceeds your budget, route hard agentic tasks to Opus 4.8 at medium effort and reserve Sonnet 5 for terminal-heavy CLI workflows where it wins on quality

The Apps Gateway Play

The gateway launch on the same day as Sonnet 5 is not coincidental. Anthropic is solving the enterprise objection to Claude Code adoption — “we cannot let developers use an AI tool that routes data through a third party” — at the exact moment they are making the default model more expensive to run. The gateway keeps inference inside your VPC. It also gives you the spend caps and telemetry you need to actually measure the tokenizer inflation before it hits your bill.

If you are evaluating Claude Code for team-wide deployment, the gateway is the piece that makes it viable. But do not confuse infrastructure readiness with cost predictability. The gateway tells you how much each developer is spending. It does not prevent the underlying cost from rising on September 1.

The Take

This is the cleanest cost trap Anthropic has shipped. The introductory discount is mathematically calibrated to keep your bills flat while the tokenizer silently inflates consumption — then standard pricing arrives after most teams have already committed their tooling. The pattern is not new, but the scale is: every Free and Pro user is now on Sonnet 5 by default, and the teams running the heaviest workloads are the ones least likely to notice a 35% effective increase until the September invoice arrives.

Your routing policy should be specific. Sonnet 5 at default effort for Terminal-Bench-style CLI workflows — it wins there outright. Opus 4.8 at medium effort for the hardest agentic coding tasks where Sonnet 5 at xhigh costs more and still scores lower. And whatever you do, run the checklist above this month. Not next month. This month. The eight-week window is not a courtesy period — it is the trap’s timer.