Claude 5.5 Family Deep Dive: Three Models in Fifteen Days, Two Leaderboards in Disagreement

When I profiled Claude Fable 5 in July, the top of the board was one model with about a point's lead. This quarter the top of the board is a dispute. The dispute is the most useful thing about the moment, because it forces a question a single-king leaderboard lets you skip: measured by whom, on what date, at what price?

Three Ships in Fifteen Days

| Model | Shipped | Anthropic's official claim | What the claim doesn't include | |-------|---------|----------------------------|--------------------------------| | Claude Opus 5.5 | Sept 22, 2026 | Matches Fable 5.1 on most work; runs at 40% lower cost than Opus 5 | Any absolute price, any benchmark score | | Claude Sonnet 5.5 | Sept 28, 2026 | 30% faster than Sonnet 5; cheaper than Sonnet 5 on most work | Any tokens-per-second figure, any price | | Claude Haiku 5.5 | Oct 7, 2026 | "The fastest and cheapest small model" (Anthropic's wording) | Any number at all |

Every claim in that table is Anthropic's own marketing math, relayed from the Anthropic Newsroom announcements with dates attached. None of it is an independent measurement. A 40% discount is a percentage in search of a denominator; I have not seen Opus 5's rate card, so I cannot tell you what the 40% comes off. "Most work" is the other load-bearing phrase: Anthropic's own claim concedes that some work does not reach the Fable 5.1 bar, and does not say which work. Read that phrase twice before you build a migration plan on it.

What can be said independently is thin but real, and it starts with the two boards.

Two Boards, Two Kings, No Reconciliation

The September 30 snapshot of the Artificial Analysis Intelligence Index v4.3, as mirrored by DataLearner, ranks Claude Opus 5.5 first. The snapshot carries the position without a published score, so the margin over the field is unknown. Nine days later, the October 9 leaderboard snapshot on llm-stats ranks GPT-5.6 Sol first at 0.590, with the tier family priced at $5/$30 for Sol, $2/$12 for Terra, and $0.20/$1.20 for Luna. The two boards also speak different scales: the v4.3 index ran around 60 at the July snapshot, while the llm-stats composite runs from 0 to 1.

A leaderboard position is a photograph. Two photographs, taken nine days apart with two cameras, can disagree without either lying. My rule this quarter is to cite every position with its tracker and its date, and refuse to average them. If you want the machinery behind that rule, the methodology breakdown covers it; the whole reshuffle gets the quarterly treatment in the Q4 model landscape.

So which board do you trust? Neither, exclusively. Trust the one whose scope matches how you will use the model, then re-check it at renewal.

Opus 5.5: The Flagship Discount

The launch claim is Fable 5.1-level performance on most work, at 40% lower running cost than Opus 5. The yardstick matters as much as the discount. Fable-tier output was, at the July snapshot, the steepest rate in Anthropic's lineup at $10 in / $50 out per million tokens (Fable 5, per our July deep dive; no rate card for the current Fable 5.1 appears in this review's snapshot). If Opus 5.5 really delivers most of that level well below Opus 5's rate — and the denominator for the 40% is still unpublished — the middle of the price-performance curve moves. The independent evidence so far is the September 30 first place on v4.3. A position is a start; the margin behind it is unpublished.

Here's the trap. The claim's omissions are doing the work a spec sheet usually does. No per-token price, no context window, no output speed, no benchmark table, and no definition of "most work." If your workload punishes slow output or tight context, you are being asked to buy on faith in a percentage.

A percentage needs a denominator.

Sonnet 5.5: The Volume Play

The claim: 30% faster than Sonnet 5, and cheaper than Sonnet 5 on most work. Sonnet-class traffic is the middle of everyone's curve, the drafting, summarizing, and coding-assist calls that never make a leaderboard but make most of the bill. Two thirties on that tier move more real money than any flagship point does, which is presumably why Anthropic led with them.

The missing numbers are the same list as above. A "30% faster" claim against an unpublished baseline is unfalsifiable from the outside, and no tokens-per-second figure appears in the announcements this review draws from. The sensible move is boring: run your own traffic through both tiers, measure cost per completed task, and let the thirties prove themselves on your bill rather than on a slide.

Haiku 5.5: The Boldest Claim, the Fewest Numbers

Anthropic's wording makes Haiku 5.5 the fastest and cheapest small model, a sentence with two superlatives and, as of this snapshot, zero published numbers behind either. A superlative without a benchmark table is a press release wearing a lab coat.

The absence has consequences specific to this tier. Small models are exactly where quality cliffs bite hardest: classification and extraction survive a cliff, multi-step reasoning does not. Neither of the two snapshots this review draws from carries a score for Haiku 5.5. No score in either snapshot is not evidence of a weak model, though procurement departments will spend it as one. Until a score shows up, the cautious assignment for this tier is high-volume, low-stakes lanes.

What Nobody Has Published

The family-wide gaps, enumerated, as of October 11:

• Per-token prices: relative claims only (40%, 30%, 30%); no dollar figures for any of the three in the announcements cited here.
• Benchmark scores: a board position for Opus 5.5 (September 30) without a published score; no appearance at all for Sonnet 5.5 or Haiku 5.5 in the two snapshots cited.
• Context windows: undisclosed for all three.
• Output speeds: undisclosed, apart from the relative "30% faster" claim.
• Availability limits: undisclosed.

Relative claims can still be true. They are just unauditable, and unauditable claims do not survive contact with a procurement department. Treat the family, for now, as a promise with percentages attached.

The Cheap End Stopped Standing Still

The discounts land in a market where the reference price has already moved. From the October 9 llm-stats snapshot: GPT-5.6 Luna at $0.20 in / $1.20 out per million tokens, Terra at $2/$12, Sol at $5/$30. From the open-source side, per a September 22 report in China's state-affiliated Economic Information Net: Xiaomi's MiMo-V2.6 topped Artificial Analysis's open-source board at 46 points, first among open models and first for a Chinese lab, at a reported $0.13 per task, passing Kimi K3 and Qwen3.8 Max. Its Pro variant sits top-three among open models on Code Arena, level with a high-spec Claude Fable 5 build. The same board lists GLM-5.3-Flash (Zhipu, 320B parameters, $0.15/$0.50), Gemini 3.7 Flash ($0.75/$3.75), Grok 4.5 from xAI ($2/$6), MiniMax M2.7 (205K context, $0.30/$1.20), and Inkling-Small from Thinking Machines Lab (276B parameters). The open-source side gets its own treatment in the open-source roundup.

Against that board, "30% cheaper than our previous model" is a conversation Anthropic is having with itself. Twenty cents per million input tokens exists one product family over. Percentages quoted against your own past answer your old invoice; they do not answer a competitor's price card.

Before You Ask Whether x-rush Has It

Here's where I'll be straight with you: x-rush.ai does not offer the Claude 5.5 family by name, and nothing in this review is an availability claim. The evaluation above is third-party editorial, the same as it would be for any lab I do not work for.

What carries over is the platform's shape. You describe the job, and each task gets routed to a model that fits it, with provider names kept out of the product vocabulary. Text generation bills a flat 0.5 credits per unit regardless of which engine served the request. If the 40% and 30% claims hold, a router is how discounts like these reach you without a migration project: the launch cycle becomes the platform's quarterly problem, and your invoice stops caring when the board reshuffles. AI Chat on x-rush is the shortest tour of that idea.

Buy, Wait, or Skip

• Opus 5.5: trial it if you were paying last-generation flagship rates for hard reasoning and the 40% claim survives contact with your workload. Wait if your sign-off process needs a price card; as of October 11, the announcements do not include one.
• Sonnet 5.5: the migration candidate for volume lanes, pending your own cost-per-task measurement. The two thirties are plausible. Plausibility is not a pilot.
• Haiku 5.5: classification, extraction, and high-volume routing; skip it for multi-step reasoning until a score exists anywhere.
• All three: skip until the price cards and benchmark tables land if your purchasing process requires auditable numbers. Given the gaps listed above, that describes more teams than it doesn't.

Fifteen days, three tiers, zero price cards. The Claude 5.5 family is either the best-value generation Anthropic has shipped or the least verifiable one, and on October 11 the evidence on hand supports both readings. Percentages age faster than price cards; this family shipped with the former and none of the latter. Re-check the board before you renew anything.

Buy the tier the task deserves. Or route each task as it comes, and keep the launch cycle off your invoice.