The small end of the 5.5 generation, completing the line alongside Opus 5.5 and Sonnet 5.5, and pitched at high-volume cost-sensitive work — summaries, compactions, database queries, classification — plus use as a subagent under the larger two. Model ID claude-haiku-5-5. Anthropic says it runs roughly 75% cheaper than Haiku 4.5 on average, and shipped alongside a halving of Sonnet 5.5's cache-read price (about 20% cheaper on most agentic work) and a monthly API credit for Max and Team subscribers.

It is the first Haiku-class model with an adjustable effort setting, so the cost-versus-intelligence trade that the larger models expose is now available at the bottom of the range. The generational jump is the striking part of the published table. Against Haiku 4.5 it goes 735 to 1620 on GDPval-AA v2.1, 614 to 1578 on AA-Briefcase v1.1, 15.7% to 72.4% on OSWorld 2.1, and 10.2% to 45.9% on Humanity's Last Exam without tools (57.4% with). On Terminal-Bench 4.0 the predecessor scores 0.0% and Haiku 5.5 reaches 39.2% — a capability that simply did not exist at this tier one generation ago.

Against competitors in its class it is ahead of GPT-6 Luna on every published row: 1620 against 1437 on GDPval-AA, 72.4% against 48.9% on OSWorld, 39.2% against 16.4% on Terminal-Bench 4.0, 46.4% against 29.1% on Chartography. The honest framing is that it remains a small model and the table says so — Sonnet 5.5 leads it on all of them (83.9% OSWorld, 70.6% Terminal-Bench, 56.9% HLE), so the case for Haiku is cost and speed at a capability level that has become usable, not parity. FrontierCode 1.1 is the closest it comes, 46.4% against Sonnet's 52.1% at xhigh effort.

Artificial Analysis scored it a week after launch: 43 on AA Intelligence Index v4.3.2 at max effort, anchored to that page per the mode rule. On v5.0 (October 27) it reads 36 at max. The new adjustable effort setting shows up directly in the index — 34 at xhigh, 30 high, 25 medium, 19 low on v5.0 (41/38/34/29 on v4.3.2), a 17-point spread on a single checkpoint, which is the clearest published illustration yet that an AA score describes a configuration and not just a model. On v4.3.2 its 43 sat below Muse Spark 1.3's 48 and well below the frontier, as it should at this tier; the result is the price it reaches it at, $0.10 / $0.50 per 1M, one to two orders of magnitude under the models it is now within ten points of.

Model Details

Context window 1,000,000
AA Intelligence 36 was 43 on v4.3.2

Benchmark Scores

Benchmark Score Mode
GDPval-AA v2.1 1620 —
AA-Briefcase v1.1 1578 —
OSWorld 2.1 (offline subset) 72.4% —
Humanity's Last Exam (no tools) 45.9% —
Humanity's Last Exam (with tools) 57.4% —
Terminal-Bench 4.0 39.2% —
FrontierCode 1.1 (Main) 46.4% —
Chartography (no tools) 46.4% —
reasoningagenticcodecomputer-useefficiency

Related