AI agents · OpenClaw · self-hosting · automation

Quick Answer

Claude Opus vs Sonnet vs Haiku: Which Model in 2026?

Published:

The short answer

Opus 5.5 for long agentic coding and knowledge work, Sonnet 5 for everyday interactive tasks, Haiku 4.5 for high-volume cheap calls, and Fable 5.1 only when you need the flagship’s max-effort reliability. That is the September 2026 answer; the tier names are stable but the models behind them change every few months, so this guide anchors on the current line-up and the durable rules for choosing between tiers.

The current line-up (September 2026)

Claude Fable 5.1Claude Opus 5.5Claude Sonnet 5Claude Haiku 4.5
RoleFlagshipLong-running agents, coding, knowledge workFast general-purposeFastest, cheapest
ReleasedSeptember 3, 2026September 22, 2026GA June 30, 20262025
Price (in / out per MTok)$10 / $50$4 / $20$2 / $10$1 / $5
Cache read$0.25 (2.5%)$0.20 (5%)$0.20 (10%)$0.10 (10%)
Batch API50% off50% off50% off50% off
Context / max output1M / 128K1M / 128K1M / 128K200K / 64K
Latency (Anthropic’s label)SlowerModerateFastFastest
ThinkingAdaptive, always onAdaptive, always onAdaptiveExtended (manual budget)
Default efforthighmediumhighNo effort parameter
Knowledge cutoffJune 2026June 2026January 2026February 2025
Subscription accessMax, Team PremiumPro and upFree and upFree and up
Next versionSonnet 5.5 (weeks)Haiku 5.5 (weeks)

Model IDs: claude-fable-5-1, claude-opus-5-5, claude-sonnet-5, claude-haiku-4-5. All four are on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry.

How the tiers differ in practice

Fable is the model Anthropic ships with its strictest safeguards and its highest default effort. Since Opus 5.5, it no longer leads most benchmarks; Anthropic’s own table shows Opus 5.5 ahead of Fable 5.1 on every row, and Artificial Analysis scores them 58 and 53. Its remaining edge is behaviour at max effort: in independent testing Fable 5.1 produced its best results at max while Opus 5.5 at max twice exhausted its 128K output budget while still reasoning. If you run single-shot, very hard problems at max, Fable is the safer default.

Opus is the workhorse. Opus 5.5 leads agentic coding (Terminal-Bench 4.0 66.4%), knowledge work (GDPval-AA v2.1 1846 Elo) and computer use (OSWorld 2.0 81.8% partial), generates output 30% faster than Opus 5, and its $0.20 cache reads make multi-hour sessions cheap. Four of its five effort levels sit on Artificial Analysis’s intelligence-versus-cost frontier. Its default effort is medium, which is where the 40% cost saving versus Opus 5 lives.

Sonnet is the tier most applications should start with. Same 1M context and 128K output as Opus, half the price, lower latency, and a permanent $2/$10 price after Anthropic cancelled the planned September 2026 rise. What you give up: a January 2026 knowledge cutoff (five months behind Opus 5.5), and lower scores on long-horizon agent benchmarks. Anthropic’s Sonnet 5.5, promised for the coming weeks, is expected to carry over Opus 5.5’s efficiency and writing improvements.

Haiku is for volume. Haiku 4.5 is the fastest model, costs $1/$5, and works well as a sub-agent, classifier or router. It is also the most dated: 200K context, 64K output, February 2025 cutoff, no adaptive thinking. Competitively it is under pressure from GPT-6 Luna at $0.10/$0.50; Haiku 5.5 is the response to watch.

Routing table

TaskModelEffortWhy
Multi-hour coding agent, migrations, auditsOpus 5.5medium–highTerminal-Bench 4.0 66.4%; 200K-line audit in <3h
Claude Code daily driverOpus 5.5mediumDefault; GitHub measured fewest steps and tokens
Interactive editing, PR-sized changesSonnet 5high (default)Half the price, lower latency, same context
Reports, financial models, decksOpus 5.5mediumGDPval leader; 16/18 fact-checked reports passed
Customer-facing chatSonnet 5mediumFast, cheap, 1M context for long threads
Classification, extraction, routing, moderationHaiku 4.5$1/$5, fastest; batch for 50% off
Sub-agents inside an Opus workflowHaiku 4.5 or Sonnet 5— / lowKeep the expensive model for planning
Hardest single-shot reasoning at maxFable 5.1maxDid not over-think to failure
Whole-repository prompts (>272K tokens)Opus 5.5 or Sonnet 5No long-context surcharge, unlike GPT-6
Offensive security, biology R&DWhichever your verification tier allowsFable and Opus 5.5 gate these behind programs

Durable rules for choosing a tier

  1. Route by task length, not task difficulty. A hard question answered in one turn is a Sonnet job; a medium-difficulty task that takes 200 tool calls is an Opus job, because per-step reliability compounds.
  2. Price the cached tokens, not the list price. On agent loops 90%+ of input is cached. Opus 5.5’s $0.20 cache read is the same as Sonnet 5’s, so the real gap between them on long sessions is closer to 2x on output than the headline suggests.
  3. Check the knowledge cutoff for API-heavy coding. A model that predates the library version you use will hallucinate signatures. Sonnet 5’s January 2026 cutoff is the one to watch here.
  4. Start one tier down and escalate on failure. Sonnet with an Opus fallback triggered by failed tests or low confidence usually beats Opus everywhere at a fraction of the cost.
  5. Re-run your effort sweep on every new model. Defaults change (Opus 5 defaulted to high, Opus 5.5 to medium) and tokens-per-effort change with them.
  6. Expect the tiers to shift. Anthropic has cut Opus prices once (Opus 5.5) and made a Sonnet increase disappear in 2026; Sonnet 5.5 and Haiku 5.5 will reshuffle this table within weeks. Anchor decisions on measured cost per task on your own workload.

For the latest Opus specifically, see What is Claude Opus 5.5?; for how the tiers compare with OpenAI’s, see GPT-6 Sol vs Luna vs Astra.

Last verified: September 23, 2026.

Sources