# Claude models explained: a Claude model comparison of Claude Opus vs Sonnet vs Haiku vs Fable — all Claude API models, Anthropic model names and prices (2026) Anthropic ships three tiers of Claude AI models — Opus, Sonnet, Haiku — plus the Fable line above them, and each generation keeps the tier names while changing the price and capability behind them. This Claude model comparison covers all Claude API models you can actually call in 2026, by job rather than by leaderboard: the Opus models for hard reasoning, Sonnet for production, Haiku for volume, and the exact Anthropic model names, API ids and list prices for each. It is written for developers choosing a model, not for benchmark-watching — the large language models below are compared on cost per task, context windows and what they are actually good at. ## The current Claude lineup at a glance List prices are Anthropic's per million tokens (input / output). On this gateway the same models bill at these list rates against credits bought below face value, so the effective cost is lower — see the live rates on each model page. Claude Fable 5.1 (id `claude-fable-5-1`) — the newest flagship: $10 / $50, 1M-token context. Strongest coding and long-horizon agent performance in the family. Claude Opus 5 (`claude-opus-5`) — $5 / $25, 1M context, with an extended-thinking variant (`claude-opus-5-thinking`). Deep reasoning at half the Fable price. Claude Sonnet 5 (`claude-sonnet-5`) — $2 / $10, 1M context. The production default and the most-called model on this gateway by a wide margin. Claude Fable 5 (`claude-fable-5`) — $10 / $50, 1M context. The previous Fable release; still served. Claude Opus 4.8 (`claude-opus-4-8`) — $5 / $25, with effort control and a `(1M)` long-context variant. The previous Opus generation. Claude Sonnet 4.6 (`claude-sonnet-4-6`) — $3 / $15, 200K with a 1M beta. Superseded by Sonnet 5 at a lower price. Claude Haiku 4.5 (`claude-haiku-4-5`) — $1 / $5, 200K. The fast, cheap tier for high-volume calls. ## Claude Sonnet vs Opus vs Haiku at a glance The short version of the Claude model comparison, before the detail. Sonnet is the default: the best price-to-capability ratio for coding and agents, and the model most production traffic on this gateway runs on. The Opus models are the flagship model tier — Opus 5 today, Opus 4.8 before it — for architecture decisions, hard debugging and long autonomous runs where a wrong step costs more than the tokens. Haiku is the small, fast tier for classification, extraction, routing and high-volume calls where the answer is short. Fable sits above Opus at the top of the range. Read as tiers rather than versions: every Anthropic model name has a concrete id behind it (`claude-sonnet-5`, `claude-opus-5`, `claude-haiku-4-5`), and the id is what the API bills. The lineup section below lists every current id with its list price; the sections after it explain how to choose between them. ## Anthropic model names and ids, explained The names people search for — "Claude Opus", "Sonnet", "Haiku" — are tiers, not versions. The API needs a concrete id such as `claude-sonnet-5` or `claude-opus-4-8`; dated aliases like `claude-sonnet-4-5-20250929` pin a snapshot. Suffixes matter: `-thinking` enables extended thinking, and a `(1M)` variant selects the long-context tier where it is priced separately. Because the tier names persist across generations, a request for "Opus" without a version can mean two very different prices. Always write the full id, and check `GET /v1/models` on the endpoint you use — the list on this gateway currently has 28 Claude ids. ## Opus vs Sonnet vs Haiku for coding Sonnet 5 is the right default for day-to-day code generation, refactoring and test writing: it is markedly cheaper than Opus while staying strong at code, and it is what Claude Code uses unless told otherwise. Move up to Opus 5 or Fable 5.1 for architecture decisions, hard debugging and long autonomous agent runs — the cases where one wrong step costs more than the token difference. Haiku 4.5 is not for the main edit loop. It is for classification, routing, extraction, summarisation and subagent scouting — anything high-volume where a small model's judgement is good enough and a big model's price is not. ## Claude Haiku vs Sonnet vs Opus: pick by job, not by rank Claude Haiku vs Sonnet vs Opus is not a question with one answer — it depends on what the call does. For a chat assistant that answers short questions, Haiku's speed and price win and the quality gap rarely shows. For code generation, refactoring, test writing and multi-step agents, Sonnet is where most teams settle: it is markedly cheaper than Opus and strong enough that Opus only pays off on the hardest steps. For long-horizon agentic work — an agent that runs for an hour across a codebase — the Opus models and Fable make fewer wrong turns, and fewer wrong turns means fewer retries, which is where their higher per-token price is recovered. A practical routing pattern: default every request to Sonnet; route classification, summarisation and subagent scouting down to Haiku; escalate to Opus only on the specific prompts where Sonnet's answers fail your own evaluation. Measured on real tasks, this usually costs less than running one model everywhere, at either end of the range. Effort levels add a second axis on the newer Opus models: the `effort` parameter (low, medium, high) trades thinking depth for latency and cost inside the same model id, so "Opus at low effort" is often a better fit than "Sonnet with a longer prompt" for reasoning-heavy but latency-sensitive work. ## Context windows: 200K vs 1M The 5-series and Fable models carry a 1M-token window natively; Opus 4.8 and Sonnet 4.6 reach it through a separate variant. A large window is free until you fill it: you pay for the tokens you send, and agentic tools resend the whole conversation every turn. Whole-repository analysis and multi-document reasoning are the honest use cases; filling 1M on every call is the most common source of a surprising bill. ## Cost per task, not cost per token Per-token price is the wrong comparison for most work. A cheaper model that needs three attempts costs more than an expensive one that gets it in one, and output tokens are priced five times input across the whole family. Three levers move the real number: effort control on Opus 4.8 (spend less reasoning on easy steps), prompt caching (cached reads are a fraction of the input price on every model), and the Batch API for anything that can wait — half price across the board. The practical rule: start one tier lower than instinct says, measure quality on your own tasks, and escalate only where it fails. ## What to know about Fable Fable is Anthropic's top line above Opus: the same API, the same tiers of context, priced at $10 / $50. Fable 5.1 is the current release; Fable 5 remains available. Reach for it when Opus 5 falls short on the hardest agentic or reasoning work, and not before — at twice the Opus price it is the last model you try, not the first. ## Models you will read about but cannot call here Anthropic has announced models beyond the public API. Claude Mythos Preview — the Mythos-class model introduced with Project Glasswing in April 2026 — is available only to selected partners for defensive security work and is not part of the public model list, on Anthropic's API or on this gateway. Older ids (Claude 3.x, Opus 4.1, Sonnet 4.5) are retired upstream and are not served either. If a model is not in `GET /v1/models` on this gateway, it cannot be called through it, whatever a blog post says. That is also the honest answer to "which Claude model is best": the best one you can actually call, priced for the job — for coding and enterprise work that is Sonnet 5 by default and Opus 5 or Fable 5.1 when the task earns it; for knowledge work and chat, Sonnet; for volume, Haiku 4.5. ## FAQ ### What are the current Claude model names? Fable 5.1, Opus 5, Sonnet 5, Fable 5, Opus 4.8, Sonnet 4.6 and Haiku 4.5. Use the API ids (claude-fable-5-1, claude-opus-5, claude-sonnet-5, and so on); -thinking and (1M) suffixes select extended thinking and long context. ### Which Claude model should developers use first? Sonnet 5. It is the production default for coding and most workloads, at $2 / $10 per million tokens list — a third of Opus. Escalate to Opus 5 only where quality actually fails. ### Claude Haiku vs Sonnet vs Opus — which is cheapest? Haiku 4.5 at $1 / $5, then Sonnet 5 at $2 / $10, then Opus 5 and Opus 4.8 at $5 / $25, then Fable at $10 / $50 (list prices per million tokens). Cheapest per token is not cheapest per task if it needs more attempts. ### Is Claude Opus 4.8 still available? Yes — served alongside Opus 5 at the same $5 / $25 list price, with effort control and a (1M) variant. Opus 5 is the newer model; keep 4.8 only where you have tuned prompts around it. ### Is AI Prime Tech an official Anthropic service? No. It is an independent gateway serving the same models through the same API format; model names are Anthropic's trademarks. ### Claude Opus vs Sonnet vs Haiku — what is the actual difference? Capability and price per token, in that order: Opus (and Fable above it) reason best and cost most; Sonnet is close on code at a fraction of the price; Haiku is fast and cheap for short, high-volume calls. Context windows are 1M on the 5-series and Fable, 200K (1M variant) on Opus 4.8 and Sonnet 4.6. ### What are the Opus models? The flagship tier of Claude AI models: Claude Opus 5 (`claude-opus-5`, with a `-thinking` variant) and the previous Claude Opus 4.8 (`claude-opus-4-8`). Both are served here at list rates against discounted credits. --- Source: https://aiprimetech.io/learn/claude-api-models-compared/ Provider: AI Prime Tech — independent Anthropic-compatible API gateway (not affiliated with Anthropic).