Choisir un modèle Claude
Débutant
Anthropic propose une famille de modèles à différents points de capacité/coût/vitesse. Bien choisir consiste surtout à faire correspondre le modèle à la tâche — et à ne pas surpayer une capacité dont vous n'avez pas besoin.
- Lire l'échelle Haiku → Sonnet → Opus comme un compromis capacité/coût/vitesse
- Partir du bon défaut au lieu de deviner, puis monter ou descendre délibérément
- Mélanger les paliers dans un seul système — le plus grand levier de coût que la plupart des gens ne tirent jamais
- Rechercher l'identifiant de modèle exact de la bonne façon, pour que les mises à niveau restent un changement d'une ligne
Les modèles actuels
Last verified: 2026-11-02 · Official models & pricing
| Model | API model ID | Tier | Best for |
|---|---|---|---|
| Claude Fable 5.1 | claude-fable-5-1 | Mythos-classflagship (above Opus) | Released September 1, 2026. For demanding reasoning and long-horizon agentic work, or when Opus 5.5 at higher effort still falls short on your evals — Anthropic's models overview recommends starting with Opus 5.5 for most workloads and reaching for Fable 5.1 only when it falls short. Same list price as Fable 5 with cache reads at a quarter of the cost. Three breaking changes vs Fable 5: forced tool_choice (any/tool) returns 400, thinking blocks are readable only by Fable 5.1 or newer, and editing earlier turns invalidates thinking blocks. Refuses in-band (HTTP 200, stop_reason: refusal); fallback targets are Opus 5 and Opus 4.8. 30-day retention (not ZDR-eligible unless expressly authorized by Anthropic); not on Priority Tier. (Supersedes Claude Fable 5, now a legacy model.) |
| Claude Mythos 5.1 | claude-mythos-5-1 | Mythos-classverification required | Same capabilities, specs and pricing as Fable 5.1, available only to organizations verified through Anthropic's verification programs. Since October 6, 2026 that is an application, not an invitation: Project Glasswing was restructured into the Cyber Verification Program's three tiers — Defense Access (organizations, and individuals on paid plans), Red Team Access (organizations only) and Specialized Access (organizations only, reviewed with the US government, where Glasswing members landed) — each granting Opus 5.5, Sonnet 5.5 and Mythos 5.1. Apply at portal.anthropic.com/programs; decisions typically within seven business days. Does not run the conversation-binding check on thinking blocks. Customers without access should use Fable 5.1. |
| Claude Opus 5.5 | claude-opus-5-5 | Opusmost capable | Anthropic's recommended starting point for most workloads, released September 22, 2026: long-running agentic coding and knowledge work, 1M context, 128k output, June 2026 reliable knowledge cutoff. Beats the larger Fable 5.1 on every row of Anthropic's launch table at 40% of its list price. Default effort is 'medium' (shared only with Haiku 5.5 — every other model that supports effort defaults to 'high'), and at a given level it thinks more per turn than Opus 5 — re-run your effort sweep instead of carrying settings over. Four breaking changes vs Opus 5: thinking cannot be disabled at any effort level (400), forced tool_choice (any/tool) returns 400 including on token counting, thinking blocks are bound to both the model and the conversation prefix, and computer_20251124 is rejected on the Claude API and Google Cloud (still accepted on Bedrock). A fifth change returns 200: text between tool calls now arrives in progress-update thinking blocks, empty at the default display of 'omitted'. Refuses in-band (HTTP 200, stop_reason: refusal) with new 'bio' and 'reasoning_extraction' categories; server-side fallback does NOT retry 'reasoning_extraction'. Fast mode is Claude API only. Retirement no sooner than September 22, 2027. (Supersedes Claude Opus 5, now a legacy model.) |
| Claude Opus 5 | claude-opus-5 | Opusmost capable | Legacy since September 22, 2026 (superseded by Opus 5.5) but still available. Complex agentic coding and enterprise work, 1M context, 128k output. Two reasons to stay on it for now: it accepts thinking: {type: 'disabled'} at 'high' effort or below, and its thinking blocks are readable by more models than Opus 5.5's (only Fable 5.1 and Mythos 5.1 read Opus 5.5's). (Supersedes Claude Opus 4.8, now a legacy model.) |
| Claude Sonnet 5.5 | claude-sonnet-5-5 | Sonnetbalanced | The best combination of speed and intelligence, released September 28, 2026: well-scoped everyday tasks, bug fixes, multistep tool use and polished documents, 1M context, 128k output (300k on Batches with a beta header), June 2026 reliable knowledge cutoff. Exactly Sonnet 5's prices and the same tokenizer, yet Anthropic claims output 30%+ faster and up to 30% less per task — the saving is in the token and turn counts, not the rate card. Beats Opus 5.5 on Terminal-Bench 4.0 (70.6% vs 66.4%) and ties it on knowledge-work Elo, but trails it by ~8 points on FrontierCode 1.1. Default effort is 'high' and levels are recalibrated — re-run your sweep. Minimum cacheable prompt drops to 512 tokens (1,024 on Sonnet 5). Five breaking changes vs Sonnet 5: thinking: {type: 'disabled'} returns 400 (use 'between_tools', accepted only at 'high' effort or below and with no companion field), forced tool_choice (any/tool) returns 400 including on token counting, thinking blocks are bound to the model, the conversation prefix and the producing account, computer_20251124 is rejected on the Claude API and Google Cloud (still accepted on Bedrock), and Opus 4.8 / Opus 4.7 / Sonnet 5 are rejected as advisor-tool advisors. A sixth change returns 200: text between tool calls arrives in progress-update thinking blocks, empty at the default display of 'omitted'. Refuses in-band (HTTP 200, stop_reason: refusal) in five categories; server-side fallback retries only 'cyber' and 'frontier_llm', on Sonnet 5. temperature/top_p/top_k at non-default values return 400. Retirement no sooner than September 28, 2027. (Supersedes Claude Sonnet 5, now a legacy model.) |
| Claude Sonnet 5 | claude-sonnet-5 | Sonnetbalanced | Legacy since September 28, 2026 (superseded by Sonnet 5.5) but still available. The balanced pick when speed and cost matter more than Opus-level depth, 1M context. Two reasons to stay on it for now: it accepts thinking: {type: 'disabled'} at any effort level including 'xhigh' and 'max' (Sonnet 5.5's 'between_tools' tops out at 'high'), and it supports forced tool_choice, which Sonnet 5.5 does not. It is also the server-side fallback target for Sonnet 5.5's 'cyber' and 'frontier_llm' refusals. (Supersedes Claude Sonnet 4.6, now a legacy model.) |
| Claude Haiku 5.5 | claude-haiku-5-5 | Haikufastest & cheapest | The fastest model, released October 7, 2026: high-volume, latency-sensitive work — classification, extraction, routing, and cheap sub-agents. 1M context (up from Haiku 4.5's 200k), 128k output (up from 64k, 300k on Batches with a beta header), June 2026 reliable knowledge cutoff (up from Feb 2025), and the first Haiku with adaptive thinking and the effort parameter (default 'medium'). THE ONLY CURRENT MODEL PRICED BY PROMPT LENGTH: every other 4.6-and-later model bills the full 1M window at standard rates, but Haiku 5.5 has two rate cards split at 100,000 tokens, and crossing that line reprices the WHOLE request — input, output, cache writes, cache reads and batch — at 5x. It also moves to the newer tokenizer, so the same text counts ~30% more tokens than on Haiku 4.5: the real saving is ~87% below the cliff and ~35% above it, not the 90% on the rate card. Five breaking changes vs Haiku 4.5: budget_tokens thinking returns 400, non-default temperature/top_p/top_k return 400, assistant prefill returns 400 even with thinking off, computer_20250124 is rejected on the Claude API and Google Cloud, and sending thinking blocks back after editing earlier turns returns 400. Unusually permissive where the bigger 5.5 models are not — it accepts both thinking: {type: 'disabled'} (at 'high' effort or below) and forced tool_choice, which Opus 5.5 and Sonnet 5.5 both reject. Refuses in-band (HTTP 200, stop_reason: refusal) in four categories with NO server-side fallback. Priority Tier is not supported. Retirement no sooner than October 7, 2027. (Supersedes Claude Haiku 4.5, now a legacy model.) |
| Claude Haiku 4.5 | claude-haiku-4-5-20251001 | Haikufastest & cheapest | Legacy since October 7, 2026 (superseded by Haiku 5.5) but still available, and still the only model with a 200k context window (not 1M), a 64k output cap, and no support for the effort parameter; its reliable knowledge cutoff is Feb 2025. Three reasons to stay on it for now: it accepts assistant prefill and manual extended thinking (budget_tokens), both of which Haiku 5.5 rejects with a 400, and it supports Priority Tier, which Haiku 5.5 does not. It is also flat-rated — Haiku 5.5's cheaper price applies only to prompts up to 100,000 tokens. Watch its lifecycle closely: its earliest-possible retirement date of October 15, 2026 has now PASSED, so it can be deprecated at any time — yet as of November 2, 2026 the status table still lists it Active with no deprecation announced. It is the only current model past its floor (Opus 4.5 is next, on November 24, 2026) and it now has a direct successor. This is a signal to plan, not an outage: Anthropic commits to at least 60 days' notice once a deprecation is announced, so a migration started now has room. |
| Claude Fable 5 | claude-fable-5 | Mythos-classflagship (above Opus) | Legacy since September 1, 2026 (superseded by Fable 5.1) but still available. Multiday autonomous runs, first-shot on ambiguous specs, dense vision, sustained delegation to parallel subagents. GA June 9 2026. Refuses in-band (HTTP 200, stop_reason: refusal) — configure server-side or client-side fallback. Adaptive thinking only; raw thinking never returned. 30-day retention (NOT ZDR-eligible). Two reasons to stay on it for now: it supports forced tool_choice and Priority Tier (existing commitments only — new ones are no longer sold), which Fable 5.1 does not. |
| Claude Mythos 5 | claude-mythos-5 | Mythos-classverification required, legacy | Superseded by Mythos 5.1. Shares Fable 5's specs and pricing, offered separately for defensive cybersecurity workflows to organizations verified through Anthropic's verification programs — originally the invite-only Project Glasswing. Since October 6, 2026 the way in is an application to the Cyber Verification Program, and what its tiers grant is Mythos 5.1, not this model. Customers without access should use Fable 5.1. |
Model IDs are exact; tiers and 'best for' are guidance. Pricing, context-window sizes and rate limits change frequently and are intentionally NOT listed here — always read them on the official page above. This file is the single source of truth: never hard-code model facts in prose, link to this table instead.
À essayer : quel modèle convient ?
Répondez à trois questions et obtenez une recommandation de départ :
Choix du modèle
Difficulté de la tâche
Ce qui compte le plus
Volume de requêtes
Recommended
Claude Sonnet
Le choix équilibré par défaut — un raisonnement et un codage solides à une fraction du coût d'Opus. Commencez ici, et ne montez en gamme que si vous atteignez un vrai plafond de qualité.
Recherchez l'ID exact du modèle dans le tableau des modèles. Règle empirique uniquement — faites une évaluation rapide sur vos propres entrées pour en être sûr.
Le modèle mental : une échelle de capacités
- Commencez avec Sonnet. C'est la bête de somme par défaut — un raisonnement et un codage solides à un coût raisonnable. La plupart des tâches devraient débuter ici.
- Montez à Opus uniquement quand Sonnet peine et que la qualité importe plus que le coût (raisonnement difficile, agents délicats, code épineux).
- Descendez à Haiku pour le travail à fort volume, sensible à la latence ou simple (classification, extraction, routage, sous-agents bon marché).
Comment choisir concrètement
Guided walkthrough1 of 4
- C'est la bête de somme équilibrée. Commencer ailleurs revient à optimiser avant d'avoir des preuves sur votre tâche réelle.
- Ne mettez pas à niveau la charge entière. Trouvez les cas où Sonnet échoue et routez seulement ceux-là vers Opus — vous achetez la qualité sans payer pour elle partout.
- Classification, extraction, routage et sous-agents bon marché ont rarement besoin d'un modèle plus gros. Testez plutôt que de présumer.
- Utilisez Haiku pour le pré/post-traitement bon marché et Sonnet/Opus pour le cœur difficile. Ce palier de modèles est l'un des plus grands leviers de coût — voir Coût et latence.
Le palier de modèles mérite sa propre lecture : Coût et latence.
Ne choisissez pas à partir des seuls benchmarks
Les benchmarks publics sont un indice de départ, pas un verdict pour votre tâche. Lancez une petite évaluation sur une poignée de vos vraies entrées avec deux modèles — cela prend quelques minutes et vaut mieux que de deviner.
Trouver l'identifiant de modèle exact
Passez toujours l'identifiant de modèle de l'API actuel (par ex. dans votre appel messages.create). Obtenez-le depuis le tableau des modèles ci-dessus ou la page officielle des modèles — et préférez le lire depuis la configuration plutôt que de le coder en dur à de multiples endroits, afin que les mises à niveau de modèle soient un changement sur une seule ligne.
Vérifiez-vous
0/4- Haiku → Sonnet → Opus est une échelle capacité/coût/vitesse — choisissez un barreau, ne devinez pas un modèle.
- Par défaut Sonnet et expédiez ; ne montez ou ne descendez qu'avec des preuves issues de votre propre tâche.
- Mettez à niveau le sous-ensemble difficile, pas la charge entière — le routage bat les mises à niveau générales.
- Mélanger les paliers dans un seul système est l'un des plus grands leviers de coût à votre disposition.
- Les benchmarks sont un indice ; une petite évaluation sur vos vraies entrées est le verdict.
- Lisez l'identifiant de modèle depuis la configuration et cherchez-le dans le tableau des modèles vivant — ne codez jamais en dur des faits de modèle.