The Tokenizer Inflates Every Prompt Before the Cliff Hits

Anthropic’s headline is a 90% input price cut: from $1.00 per million tokens on Haiku 4.5 to $0.10 on Haiku 5.5 for short prompts. That rate applies to requests up to 100,000 tokens. Cross that threshold and every rate multiplies by five — $0.10 becomes $0.50 on input, $0.50 becomes $2.50 on output. The higher rate applies to every token in the request, not just the tokens above 100K.

The migration complication is the tokenizer. Haiku 5.5 uses a newer tokenizer that encodes the same text as roughly 30% more tokens than Haiku 4.5, according to Anthropic’s migration guide. The effective 100K threshold maps to approximately 77,000 Haiku 4.5 tokens. A code-review prompt that ran at 80,000 tokens on Haiku 4.5 will run at roughly 104,000 on Haiku 5.5 and hit the long-context tier — without any change to the request.

Anthropic says approximately 90% of Haiku 4.5 requests fall below the new threshold. The other 10% pay between 35% and 50% less than Haiku 4.5’s flat $1.00 rate — still cheaper, but not 90% cheaper. Whether cached tokens count toward the 100K threshold is not stated in Anthropic’s pricing documentation.

What the Benchmarks Actually Show

All benchmark numbers in Anthropic’s announcement are vendor-run. Haiku 5.5 scored 72.4% on OSWorld 2.1 (a desktop-operation benchmark), versus 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna. On Terminal-Bench 4.0 (command-line agent tasks), Haiku 5.5 reached 39.2% — up from 0.0% for Haiku 4.5 — while GPT-6 Luna scored 16.4%. On FrontierCode 1.1 Main (coding), Haiku 5.5 scored 46.4% versus Luna’s 42.4%. Anthropic’s GDPval-AA v2.1 Elo score for Haiku 5.5 is 1,620 versus 1,437 for Luna.

Vals.ai ran independent evaluations at maximum compute effort. On Terminal-Bench 4.0, Vals.ai scored Haiku 5.5 at 35.35% (rank 12 of 45 models), roughly four percentage points below Anthropic’s self-reported figure — a common but not negligible gap for vendor evals. On Vibe Code Bench v1.1, Vals.ai ranked Haiku 5.5 third among 110 models at 90.44%, behind Gemini 4 Argon (91.91%) and Claude Sonnet 5.5 (92.39%). On CyberBench v1.1, Haiku 5.5 placed fifth of 45 models at 75.48%.

The Vals.ai composite Vals Index placed Haiku 5.5 at 54.31%, ranking 16th among 45 models across their full benchmark suite. That positions it solidly in the mid-tier — capable enough for classification, extraction, and subagent subtasks, not suited for complex multi-step reasoning without a larger model handling planning.

Cost Math for Three Real Workloads

Prices below use Anthropic’s published list rates for Haiku 5.5 and Haiku 4.5, and OpenAI’s published list rates for GPT-6 Luna. Token counts in Haiku 5.5 terms account for the ~30% tokenizer inflation. All figures are per 10,000 requests.

Math: Haiku 4.5 at $1.00/$5.00 input/output per MTok. Haiku 5.5 short at $0.10/$0.50, long at $0.50/$2.50. GPT-6 Luna short (up to 272K tokens) at $0.10/$0.50. Sources: Anthropic and OpenAI published list rates.

The third row is the case to watch. A prompt that ran at 80K tokens on Haiku 4.5 runs at roughly 104K on Haiku 5.5 and falls into the long-context tier. Haiku 5.5 is still 35% cheaper than Haiku 4.5 there — but GPT-6 Luna, which keeps the $0.10 input rate up to 272K tokens, comes in at roughly one-sixth the Haiku 5.5 long-tier cost for the same call.

How the Three Models Price at Each Context Length

* Anthropic-run benchmarks (vendor-run): OSWorld 2.1 72.4% (Haiku 5.5) vs 48.9% (GPT-6 Luna); Terminal-Bench 4.0 39.2% vs 16.4% (Anthropic). GPT-6 Luna pricing from OpenAI published list rates.

For Swiss & EU teams

Haiku 5.5 is available on Amazon Bedrock and Google Cloud Vertex AI, both of which offer EU data-residency regions. Anthropic’s launch announcement does not mention EU-specific data processing terms for the new model version. Teams subject to FADP or GDPR should verify that their existing Bedrock or Vertex data processing agreements cover Haiku 5.5 before routing live data through it. Anthropic has published no statement on EU AI Act classification for Haiku 5.5 as of its 7 October 2026 release date.

Six API Changes That Return HTTP 400

Haiku 5.5 is stricter about inference parameters than Haiku 4.5. These six changes break existing code and require fixes before migration:

• temperature must be exactly 1. Any other value returns a 400 error. Haiku 4.5 accepted the full 0–1 range.
• top_p must be exactly 0.99 or omitted. Any other value returns 400.
• top_k is rejected entirely. Remove it from all requests.
• Assistant prefill on a final assistant turn is rejected. A trailing assistant message in the conversation array now returns 400.
• budget_tokens syntax for thinking is rejected. Migrate to the thinking block format with the effort parameter instead.
• The older computer-use tool schema returns 400 on both the Claude API and Google Cloud. Use the updated tool definition from Anthropic’s migration guide.

The model also introduces a new stop_reason value of “refusal” with no server-side fallback. Code that enumerates stop reasons without a default branch will need updating. Priority Tier routing is not supported on Haiku 5.5.

What to Run Before You Deploy

The 90% price cut is real for the workloads it applies to. For any pipeline where requests stay below roughly 77K Haiku 4.5 tokens, migration cuts costs by 87–90% with the benchmark gains included. The OSWorld jump — 15.7% to 72.4% — makes Haiku 5.5 viable for lightweight computer-use and browser subagents that Haiku 4.5 handled poorly.

Before deploying: measure your 95th-percentile request size in Haiku 4.5 tokens, multiply by 1.3, and check whether it stays below 100K. Fix the six API parameters. For any workload that routinely exceeds 100K tokens, run a cost comparison against GPT-6 Luna — at that length, Luna is five times cheaper on input at identical headline pricing. See last week’s Model Scorecard for the release overview, and the guide on evaluating a new frontier model before switching for an eval framework you can run on your own workload.

Use it if your prompts average under 77K Haiku 4.5 tokens and you are not using temperature, top_k, or assistant prefill. Test it if your prompt distribution straddles the 100K line — model the cost split before committing. Skip it for sustained long-context work; route those calls to GPT-6 Luna at $0.10/M input.

Further Reading

• Claude Haiku 5.5: Pricing, Benchmarks and the 100K Rule (Roo) — Detailed breakdown of the 100K cliff, tokenizer impact, and independent cost calculations
• Claude Haiku 5.5 — Vals.ai — Independent evaluation across 22 benchmarks including Terminal-Bench, CyberBench, and Vibe Code Bench
• Claude Haiku 5.5 100K Price Step (OrcaRouter) — Analysis of which workload categories are most at risk of hitting the long-context tier

Your turn: Have you measured what the 30% tokenizer inflation does to your 95th-percentile request size? Did any prompts you expected to stay short cross the 100K line? Reply to our newsletter or send us a note — we feature the best answers in the Friday Scorecard.

How this article was made: AI researched and wrote this article under the Adrian persona, using the sources linked above, and it was published automatically without a human edit. Editorial guidelines are set by Adi. Spotted an error? Tell us and we will correct it.