{"slug": "claude-haiku-5-5-s-scorecard-stop-reading-the-rank-start-routing-tasks", "title": "Claude Haiku 5.5's scorecard: stop reading the rank, start routing tasks", "summary": "Anthropic released Claude Haiku 5.5 (model ID claude-haiku-5-5) on 7 October 2026 with a comparison table against Haiku 4.5, GPT-6 Luna and Sonnet 5.5. According to Anthropic's published figures, the model jumps from 15.7% to 72.4% on OSWorld 2.1 and from 0.0% to 39.2% on Terminal-Bench 4.0 over Haiku 4.5, while trailing Sonnet 5.5 on long-horizon agentic coding (39.2% vs 70.6%). A developer argues the scorecard should be read as a routing guide rather than a ranking, noting the listed unit price falls 90% but a newer tokenizer yields roughly 75% lower average running cost.", "body_md": "Anthropic released **Claude Haiku 5.5** (model ID `claude-haiku-5-5`) on 7 October 2026 and published a comparison table against Haiku 4.5, GPT-6 Luna and Sonnet 5.5. Most takes I've seen boil it down to \"where does it rank\". That's the least useful way to read it. The better question: **which work can I move down to the cheap tier now?**\n\nAll figures below come from Anthropic's announcement, not from an independent measurement.\n\n| Benchmark | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 | \n|---|---|---|---|---|\n| GDPval-AA v2.1 (knowledge work) | 1620 | 735 | 1437 | 1840 | \n| AA-Briefcase v1.1 (knowledge work) | 1578 | 614 | 1336 | 1824 | \n| OSWorld 2.1 (computer use) | 72.4% | 15.7% | 48.9% | 83.9% | \n| HLE, no tools | 45.9% | 10.2% | — | 56.9% | \n| HLE, with tools | 57.4% | 18.7% | — | 64.5% | \n| Terminal-Bench 4.0 (agentic coding) | 39.2% | 0.0% | 16.4% | 70.6% | \n| FrontierCode 1.1 (agentic coding) | 46.4% | — | 42.4% | 52.1% | \n| Chartography (visual reasoning) | 46.4% | 6.4% | 29.1% | 61.6% | \n\nThese measure different things. Collapsing them into a single \"it's #2\" hides the most important signal: the **spread** across tasks.\n\n**1. A generational jump over Haiku 4.5.** OSWorld goes from 15.7% to 72.4%, Terminal-Bench from 0.0% to 39.2%. Both are long action chains: understand, call a tool, handle feedback, recover from a mistake. This is not \"nicer wording\"; small models are no longer just for short classification jobs.\n\n**2. Ahead of GPT-6 Luna on the items reported side by side.** But it's the vendor's table and the vendor's config. Fair phrasing: *on the items Anthropic published*, Haiku 5.5 scored higher. Don't extrapolate to untested tasks.\n\n**3. Close to Sonnet 5.5 in places, but the tiers are real.** OSWorld 72.4% vs 83.9% is a gap; Terminal-Bench 39.2% vs 70.6% is a canyon. Long-horizon, multi-file, error-recovering agent work still belongs to the bigger model.\n\nUp to 100K tokens: $0.10 / MTok input, $0.50 / MTok output, $0.01 cache read. Over 100K: $0.50 / $2.50 / $0.05. Listed unit price vs Haiku 4.5 drops 90%, but a newer tokenizer uses slightly more tokens, so Anthropic quotes about 75% lower *average* running cost.\n\nThe metric that matters in production:\n\n```\ncost per successful task = total spend to finish the work / tasks finished successfully\n```\n\nIf a cheap call needs two retries, or a chunk of requests gets escalated to Sonnet anyway, the saving shrinks. If the output is auto-checkable and retries are cheap, the economics are excellent.\n\nThe 100K boundary also nudges you toward a layered design instead of stuffing everything into one prompt:\n\nHaiku 5.5 is the first Haiku with an adjustable **effort** setting. One score = one configuration. What you really want is a cost–quality curve. Low effort for classification, format conversion, fixed-field extraction; higher effort for long-document summaries, cross-paragraph synthesis, multi-step browser work. But cranking effort doesn't erase the Terminal-Bench gap.\n\n| Task | Default | \n|---|---|\n| Classification, routing, labels | Haiku 5.5 | \n| Field extraction / data cleanup | Haiku 5.5 | \n| Summaries, context compaction | Haiku 5.5 | \n| Short browser flows | Haiku 5.5 | \n| Code explanation, small fixes | Haiku or Sonnet (depends on tests + cost of mistakes) | \n| Multi-file edits | Sonnet 5.5 | \n| Long-running agentic coding | Sonnet or Opus 5.5 | \n| Final review, high-stakes calls | Sonnet or Opus 5.5 | \n\nValidate on your own tasks: can failures be detected automatically? Is a retry worth it? Does lower latency actually improve the product? Does human review eat the savings?\n\nHaiku 5.5 doesn't beat Sonnet 5.5. It's strong because it can be used **in volume**: a batch of work that used to need a mid-size model now fits in a faster, cheaper execution layer.\n\nThe full write-up (with the reading method for official benchmarks) is on Folkbench, and browsing it needs no signup: [Folkbench – Haiku 5.5 official scorecard](https://folkbench.com/en/blog/claude-haiku-5-5-official-scorecard?utm_source=luntan&utm_campaign=dev)\n\nHow are you splitting work between small and large models today?", "url": "https://wpnews.pro/news/claude-haiku-5-5-s-scorecard-stop-reading-the-rank-start-routing-tasks", "canonical_source": "https://dev.to/sichi_chen_a4a87b20aa2dbf/claude-haiku-55s-scorecard-stop-reading-the-rank-start-routing-tasks-1m54", "published_at": "2026-10-29 07:07:43+00:00", "updated_at": "2026-10-29 07:17:09.216878+00:00", "lang": "en", "topics": ["large-language-models", "ai-products", "ai-agents", "ai-tools"], "entities": ["Anthropic", "Claude Haiku 5.5", "Claude Haiku 4.5", "Claude Sonnet 5.5", "GPT-6 Luna", "Folkbench"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/claude-haiku-5-5-s-scorecard-stop-reading-the-rank-start-routing-tasks", "markdown": "https://wpnews.pro/news/claude-haiku-5-5-s-scorecard-stop-reading-the-rank-start-routing-tasks.md", "text": "https://wpnews.pro/news/claude-haiku-5-5-s-scorecard-stop-reading-the-rank-start-routing-tasks.txt", "jsonld": "https://wpnews.pro/news/claude-haiku-5-5-s-scorecard-stop-reading-the-rank-start-routing-tasks.jsonld"}}