Claude Model Board: Fable 5.1, Opus 5, Sonnet 5, Haiku 4.5

Analysis

Start with Opus 5, not the flagship

The instinct with a four-model lineup is to reach for the most capable one. Anthropic’s own model-selection guidance says the opposite: start with Claude Opus 5 for most workloads. Fable 5.1 is positioned for demanding reasoning and long-horizon agentic work, or for cases where evals on Opus 5 at higher effort still fall short. That is a narrower trigger than "important task."

Sonnet 5 is the speed and intelligence balance, and at $2 / $10 with a 1M context window it is now materially cheaper than it looked in July, when the published plan was a September increase to $3 / $15. Haiku 4.5 remains the fastest and cheapest model for bounded work.

Analysis

Current lineup

The table below is the operational version: the identifier you put in code, what it costs, how much it can read and write, and when it can stop being available. Retirement dates are the earliest possible dates published for Anthropic platforms, not scheduled shutdowns.

• Default thinking effort is high on Fable 5.1, Opus 5, and Sonnet 5 at the platform level.
• Haiku 4.5 also carries a dated alias, claude-haiku-4-5-20251001, for pinned deployments.
• Cache reads are 10% of base input on most models and 2.5% on Fable 5.1 and Mythos 5.1.
• The Batch API applies a 50% discount for non-interactive workloads.
• Haiku 4.5 has the nearest retirement floor of the four; pin and monitor it if it is in a production path.

Analysis

Mythos 5.1 is an access program, not a model choice

Mythos 5.1 is the same underlying model as Fable 5.1 with more permissive safeguards, available only through trusted access under Project Glasswing. It prices identically at $10 / $50 with $0.25 cache reads. For public planning, treat it the way you would treat any allowlisted program: assume you do not have it unless your account proves otherwise.

The distinction matters when reading benchmark tables, because some published Mythos 5.1 figures exceed the Fable 5.1 figures on the same eval. A number you cannot reproduce on the model you can actually call is not a number you can plan with.

Analysis

Prior generation: still listed, no longer current

Fable 5, Opus 4.8, 4.7, 4.6, and 4.5, and Sonnet 4.6 and 4.5 are still listed but are not the current board. Two of these come up constantly in stale documentation: Opus 4.8 stopped being the current Opus on July 24, 2026, and Fable 5 stopped being the flagship on September 1, 2026.

If you are auditing a codebase or a runbook, those two strings are the highest-yield things to search for. Historical model timelines, deprecation archaeology, and pricing history are out of lane for this site; Claude Weekly covers the current operating picture.

Analysis

How to route without guessing

Routing by prestige produces expensive systems that are not measurably better. Routing by eval produces defensible ones. Build a small task suite that includes your actual failure modes: refusal boundaries, tool-use JSON validity, long-document recall, agentic loop length, voice, latency, and cost per successful answer.

• Start at Opus 5 and only escalate to Fable 5.1 when a measured gap justifies the higher rate.
• Escalate effort on Opus 5 before escalating models; effort is cheaper than a price tier.
• Route bounded, repetitive work to Haiku 4.5 and keep the savings visible in dashboards.
• Log model IDs on every production call so incident review can distinguish surface, model, and prompt.
• Recheck cache economics per model, not once for the whole platform.