aicharts
Anthropic released Claude Haiku 5.5 on October 7, 2026. In the Intelligence Index snapshot retrieved Oct 10, 2026, 5:51 AM UTC, Claude Haiku 5.5 (Max) scores 43.4 at $0.213 per task, 32nd of the 108 comparable configurations, meaning the rows with a measured cost per task. One cheaper row, MiMo-V2.6-Pro, scores higher, so the max row is just off the cost frontier; the xhigh row is a vertex of it. The four lower effort levels of the same model score 41.2, 37.8, 34.5, and 29.5, and the cheapest of them costs $0.024 per task. The coding-agent chart stores no Claude Haiku 5.5 row, so this note has no coding-agent score to set beside the Index.
The evaluations behind each row
The Artificial Analysis Intelligence Index runs a model through its API under one harness that is the same for every model, across 10 evaluations weighted 30% agents, 20% coding, 20% scientific reasoning, and 30% general capability, at version 4.3.2. The evaluations are AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, and AA-LCR v1.1. Cost per task is the weighted average API bill for one task across those evaluations, split into input-side cost (non-cached input, cache reads, and cache writes) and output-side cost (reasoning and answer tokens).
A row on the chart is one configuration: the model at one effort level with the settings Artificial Analysis names in the row. The Claude Haiku 5.5 rows carry nothing in their names but the effort level, max, xhigh, high, medium, and low, so every difference between them in this note is a difference in effort.
Ranked 32nd of 108, one row off the cost frontier
The 43.4 places Claude Haiku 5.5 (Max) 32nd of the 108 comparable configurations in the snapshot retrieved Oct 10, 2026, 5:51 AM UTC. Of the rest, 31 configurations score higher; the leader, Claude Opus 5.5 (Max, Default Fallback), is 14.2 points above it at $5.98 per task. The row is not on the chart’s cost frontier, and one row keeps it off: MiMo-V2.6-Pro scores 46.3 for $0.133 per task, 2.9 points higher for 0.6x the cost. It is also the cheapest configuration that scores higher.
Of the 108, 37 cost less per task than the $0.213. Within one index point of it sit two configurations, and both cost more: Step 5 Preview at 43.7 for $1.03 per task and Kimi K3 (Max) at 43.6 for $2.00 per task, which is 4.8x and 9.4x the Haiku row’s cost.
The table lists the five highest-scoring configurations after it, with each cost as a multiple of its $0.213. GPT-6.1 Sol (Low) gives up 1.3 points for 0.6x the cost.
| Configuration | Intelligence Index | Cost per task | Points below Claude Haiku 5.5 (max) | Multiple of its cost |
|---|---|---|---|---|
| Claude Opus 5.5 (Low, Default Fallback) | 42.3 | $0.551 | 1.1 points | 2.6x |
| Grok 4.7 (Low) | 42.2 | $1.25 | 1.2 points | 5.9x |
| GPT-6.1 Sol (Low) | 42.1 | $0.131 | 1.3 points | 0.6x |
| GPT-5.6 Terra (Max) | 42.1 | $1.40 | 1.3 points | 6.6x |
| GLM-5.3-Flash | 41.8 | $0.253 | 1.6 points | 1.2x |
One of the five Claude Haiku 5.5 effort rows is a vertex of the chart’s cost frontier, the set of rows no other row beats on both score and cost. Claude Haiku 5.5 (Xhigh) at 41.2 for $0.124 per task sits below GPT-6.1 Sol (Low) at 42.1 for $0.131 per task, giving up 0.8 points for 0.9x the cost; the next vertex down, GPT-6 Luna (Max) at 38.1 for $0.068 per task, gives up 3.1 points for 0.5x the Haiku row’s cost.
What each effort step buys
The snapshot stores five comparable Claude Haiku 5.5 rows with an effort level, one per level, and no row without one. From low to max, the score moves from 29.5 to 43.4 index points and the cost per task from $0.024 to $0.213: 13.9 points for 8.7x the cost. The table lists the levels cheapest first and states what each step buys over the level below it.
| Effort level | Intelligence Index | Cost per task | Output tokens per task | Reasoning share of output tokens | Points over the level below | Cost multiple of the level below |
|---|---|---|---|---|---|---|
| low | 29.5 | $0.024 | 17,286 | 53% | - | - |
| medium | 34.5 | $0.047 | 32,813 | 60% | +5.0 | 1.9x |
| high | 37.8 | $0.079 | 55,388 | 67% | +3.4 | 1.7x |
| xhigh | 41.2 | $0.124 | 89,394 | 74% | +3.4 | 1.6x |
| max | 43.4 | $0.213 | 162,164 | 80% | +2.1 | 1.7x |
The largest step is from low to medium: +5.0 points at 1.9x the cost per task. The smallest is from xhigh to max: +2.1 points at 1.7x the cost per task. Each step multiplies the cost per task by between 1.6x and 1.9x and buys between +2.1 and +5.0 points. At max, the model writes 162,164 output tokens per task and 80% of them are reasoning tokens; the share runs from 53% to 80% across the five levels.
Anthropic’s announcement says that “Haiku 5.5 is our first Haiku-class model to come with an adjustable effort setting” and names no default level, so this note places none. The snapshot dates every level October 7, 2026.
What the $0.213 pays for
Of the $0.213 per task at max, $0.132 is input-side cost and $0.081 is output. Largest component first, the five parts are cache reads at $0.095, reasoning tokens at $0.065, cache writes at $0.034, answer tokens at $0.017, and non-cached input at $0.003. Cache reads are 45% of the total. Input-side cost is 62% of the total at max and stays between 62% and 65% across the five levels, so most of what a Haiku task costs on the Index is reading context back in, not writing the answer.
| Effort level | Input cost | Of which cache reads | Output cost | Of which reasoning tokens | Input share of total |
|---|---|---|---|---|---|
| low | $0.016 | $0.008 | $0.009 | $0.005 | 65% |
| medium | $0.031 | $0.019 | $0.016 | $0.010 | 65% |
| high | $0.052 | $0.035 | $0.028 | $0.018 | 65% |
| xhigh | $0.079 | $0.056 | $0.045 | $0.033 | 64% |
| max | $0.132 | $0.095 | $0.081 | $0.065 | 62% |
Artificial Analysis’s model page, captured October 12, 2026 UTC, lists the prices behind these figures as $0.10 per million input tokens and $0.50 per million output tokens with a 90% cache discount, and rounds the row to an index score of 43 at $0.21 per task. It records 440M output tokens to run the whole index, which it calls “very verbose in comparison to the median of 100M,” measures 235.3 output tokens per second, lists the model as proprietary with a 1M token context window, released October 7, 2026, and sums the row up as “amongst the leading models in intelligence and well priced when comparing to other models of similar price” and “notably fast, however very verbose.”
The same page ranks the row #2 / 180 on intelligence. That rank is among the models the page groups with it by price range, not across the whole index; on this chart, which ranks every configuration with a measured cost, the row is 32nd of 108. The two numbers answer different questions and neither corrects the other.
Haiku against Sonnet 5.5 and Opus 5.5 on the same chart
The snapshot stores 20 configurations from Anthropic with a measured cost per task: five Claude Opus 5.5, five Claude Sonnet 5.5, five Claude Fable 5.1, and five Claude Haiku 5.5. The table measures every Sonnet 5.5 and Opus 5.5 row against Claude Haiku 5.5 (Max), and marks whether each row is a vertex of the cost frontier drawn through Anthropic rows alone.
| Configuration | Intelligence Index | Cost per task | Points against Claude Haiku 5.5 (max) | Multiple of its cost | On the Anthropic-only frontier |
|---|---|---|---|---|---|
| Claude Opus 5.5 (Max, Default Fallback) | 57.6 | $5.98 | +14.2 | 28.1x | yes |
| Claude Sonnet 5.5 (Max, Default Fallback) | 56.0 | $5.46 | +12.6 | 25.7x | yes |
| Claude Opus 5.5 (Xhigh, Default Fallback) | 56.0 | $3.46 | +12.6 | 16.3x | yes |
| Claude Opus 5.5 (High, Default Fallback) | 53.6 | $1.82 | +10.2 | 8.6x | yes |
| Claude Sonnet 5.5 (Xhigh, Default Fallback) | 51.9 | $2.01 | +8.5 | 9.5x | no |
| Claude Opus 5.5 (Medium, Default Fallback) | 51.2 | $1.34 | +7.8 | 6.3x | yes |
| Claude Sonnet 5.5 (High, Default Fallback) | 46.8 | $0.884 | +3.4 | 4.2x | yes |
| Claude Opus 5.5 (Low, Default Fallback) | 42.3 | $0.551 | −1.1 | 2.6x | no |
| Claude Sonnet 5.5 (Medium, Default Fallback) | 40.8 | $0.483 | −2.6 | 2.3x | no |
| Claude Sonnet 5.5 (Low, Default Fallback) | 35.9 | $0.345 | −7.5 | 1.6x | no |
A Haiku row beats three configurations from the other Anthropic releases on both axes: Claude Opus 5.5 (Low, Default Fallback) scores 1.1 points below Claude Haiku 5.5 (Max) at 2.6x its cost, Claude Sonnet 5.5 (Medium, Default Fallback) scores 2.6 points below Claude Haiku 5.5 (Max) at 2.3x its cost, and Claude Sonnet 5.5 (Low, Default Fallback) scores 7.5 points below Claude Haiku 5.5 (Max) at 1.6x its cost.
Drawing the cost frontier through the Anthropic rows alone and walking it from the top, the vertices are Claude Opus 5.5 (Max, Default Fallback), Claude Sonnet 5.5 (Max, Default Fallback), Claude Opus 5.5 (Xhigh, Default Fallback), Claude Opus 5.5 (High, Default Fallback), Claude Opus 5.5 (Medium, Default Fallback), Claude Sonnet 5.5 (High, Default Fallback), Claude Haiku 5.5 (Max), Claude Haiku 5.5 (Xhigh), Claude Haiku 5.5 (High), Claude Haiku 5.5 (Medium), and Claude Haiku 5.5 (Low). Below Claude Sonnet 5.5 (High, Default Fallback) at $0.884 per task, every vertex is a Claude Haiku 5.5 row: within Anthropic’s range on this chart, the cheapest row that reaches any score up to the 43.4 of Claude Haiku 5.5 (Max) is a Haiku effort level.
Anthropic’s prices, its benchmark table, and GPT-6 Luna
Anthropic’s announcement of October 7, 2026 opens with “Introducing Claude Haiku 5.5: the cheapest, fastest, and most capable small model we’ve ever released,” and says that “Claude Haiku 5.5 is designed for high-volume, cost-sensitive tasks.” For prompts up to 100,000 tokens it prices the model at $0.10 per million input tokens, $0.50 per million output tokens, $0.01 per million cache-read tokens, and $0.125 per million cache-write tokens; over 100,000 tokens the four prices are $0.50, $2.50, $0.05, and $0.625. Against Haiku 4.5 at $1.00, $5.00, and $0.10 for input, output, and cache reads, Anthropic writes that “On average, it now costs around 75% less to run,” with a footnote that the model is “priced 90% lower than Claude Haiku 4.5 for requests up to 100,000 tokens, and 50% lower for requests over 100,000 tokens,” and calls it “especially good value when used for tasks with prompts up to 100,000 tokens, which make up around 90% of requests to our previous Haiku model.” In the snapshot’s max row, cache reads are $0.095 of the $0.213 per task, so the cache-read price is the one that moves the chart’s cost column most. The snapshot stores no Claude Haiku 4.5 row, so the 75% cannot be checked on the chart. Developers reach the model on the Claude Platform as claude-haiku-5-5.
The announcement’s benchmark table reports Anthropic’s own runs against Haiku 4.5, GPT-6 Luna, and Sonnet 5.5: GDPval-AA v2.1 at 1620, OSWorld 2.1 at 72.4% on an offline subset, Humanity’s Last Exam at 45.9% without tools, Terminal-Bench 4.0 at 39.2%, FrontierCode 1.1 (Main) at 46.4%, and Chartography at 46.4%. Those are vendor numbers; none of them appears on an aicharts chart. Anthropic adds that “Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding tasks like those measured by Terminal-Bench 4.0.” Terminal-Bench 4.0 and GDPval-AA v2.1 are also two of the 10 evaluations inside the Intelligence Index, where Artificial Analysis runs them under its own harness; the snapshot stores the composite score and not the per-evaluation results, so the 39.2% cannot be checked against it here.
GPT-6 Luna, the one model outside Anthropic’s range in that table, does have rows on the chart: six configurations with a measured cost, from GPT-6 Luna (Max) at 38.1 for $0.068 per task down to GPT-6 Luna (Non-reasoning) at 18.5 for $0.011 per task. The highest Luna row scores 5.3 points below Claude Haiku 5.5 (Max) at 0.3x its cost, and on the chart’s cost frontier it is the vertex directly below Claude Haiku 5.5 (Xhigh), 3.1 points lower for 0.5x the cost.
No Haiku row on the coding-agent chart yet
The coding-agent chart is a daily snapshot of the public Artificial Analysis coding-agents comparison. A row on it is a model running inside a named agent harness at one effort setting, and its AA Index averages three coding benchmarks: DeepSWE v1.1, Terminal-Bench 4, and SWE-Atlas-QnA. Its cost is the mean API bill for one task in that harness, including every tool call and repeated context the harness sends.
In the coding-agent snapshot retrieved Oct 5, 2026, 2:10 PM UTC, no row runs Claude Haiku 5.5 in any harness, so this note has no coding-agent score for the model and does not estimate one. The Anthropic rows it stores are Claude Code with Sonnet 5.5 at five settings, Claude Code with Opus 5.5 at max, Claude Code with Fable 5.1 (with fallback) at max, and Claude Code with Opus 5 at max. Anthropic’s announcement quotes Cognition saying that “Claude Haiku 5.5 joins the sidekick lineup in Devin Fusion as an excellent option,” and a Haiku row on the coding-agent chart would be the model inside a harness like that one; the snapshot does not store one yet.
Limits
- The chart scores, costs, and token counts are Artificial Analysis measurements of the Claude Haiku 5.5 rows on the retrieval date under Intelligence Index version 4.3.2. They say nothing about other tasks, prompts, or harnesses.
- The rank, the frontier vertices, the nearest-score table, the effort ladder, the cost shares, and the Anthropic-only frontier are computed from the snapshot by aicharts. A new, removed, or rescored configuration moves them, and the snapshot updates on a schedule.
- Claude Haiku 5.5 inside Claude Code, Devin Fusion, or another harness is a configuration the coding-agent snapshot does not store, so this note says nothing about it.
- The per-evaluation scores behind the composite are not stored, so no sentence here says which of the 10 evaluations an effort step helps most.
- The prices, the 75% saving against Haiku 4.5, and the benchmark table are Anthropic’s. aicharts did not run Claude Haiku 5.5.