Picking a Claude Model for Agent Work: Fable 5.1, Opus 5, Sonnet 5, Haiku 4.5
A September 2026 engineering reference — routing defaults, cache-read economics, and the Fable 5.1 changes that will break an existing agent harness.
This is a reference, not a build log. Every number below comes from Anthropic’s published documentation, pricing pages, and announcements, linked inline. Where I do arithmetic, it is arithmetic on their list prices and I show the work. Where I quote an eval score, it is Anthropic’s published score, not something I measured.
If you wrote an agent harness against Claude in early 2026 and have not touched the model configuration since, three things have changed enough to matter: Opus 5 shipped in July, Fable 5.1 shipped on September 1 and inverted the cache-read cost ordering, and Fable 5.1 introduced behavioral changes that will surface as runtime errors rather than as quietly worse output.
The lineup as of September 2026
| Model | API ID | In / out per MTok | Context | Max out | Default effort | Knowledge cutoff |
|---|---|---|---|---|---|---|
| Claude Fable 5.1 | claude-fable-5-1 | $10 / $50 | 1M | 128K | high | Jun 2026 |
| Claude Opus 5 | claude-opus-5 | $5 / $25 | 1M | 128K | high | May 2026 |
| Claude Sonnet 5 | claude-sonnet-5 | $2 / $10 | 1M | 128K | high | Jan 2026 |
| Claude Haiku 4.5 | claude-haiku-4-5 | $1 / $5 | 200K | 64K | n/a | Feb 2025 |
Source: the models overview and pricing.
Fable 5.1, Opus 5, and Sonnet 5 all use adaptive thinking; on Fable 5.1 it is always on. Haiku 4.5 uses extended thinking and is the only model in the list still on a 200K context window and a Feb 2025 cutoff — which is the real reason to be careful about using it for anything that needs current library APIs.
Retirement is not imminent for any of them. Anthropic’s published floors are 2027-09-01 for Fable 5.1, 2027-07-24 for Opus 5, 2027-06-30 for Sonnet 5, and 2026-10-15 for Haiku 4.5. Haiku 4.5 is the one worth putting a calendar reminder on.
Anthropic’s own routing guidance is to start with Opus 5 for most workloads, and reach for Fable 5.1 for demanding reasoning and long-horizon agentic work — or when your evals on Opus 5 at higher effort still fall short. Sonnet 5 is the speed/intelligence balance point and Haiku 4.5 is the cheap, fast tier.
Two corrections to claims that are still circulating:
- Opus 4.8 is not the current Opus. Opus 5 launched July 24, 2026 at the same $5/$25 as Opus 4.8.
- Sonnet 5 pricing did not go up on September 1. The $2/$10 rate was announced as introductory through 2026-08-31 and was widely expected to become $3/$15. It did not. $2/$10 is now the standard rate; the pricing page is the authority.
Cache reads are where the cost model stopped being intuitive
The interesting consequence of the Fable 5.1 launch is not the headline price — it is the cache-read rate.
Most Claude models bill a cache read at 10% of base input. Fable 5.1 and Mythos 5.1 bill cache reads at 2.5% of base input. Multiply that out against the list prices above:
| Model | Base input / MTok | Cache read rate | Cache read / MTok |
|---|---|---|---|
| Claude Fable 5.1 | $10 | 2.5% | $0.25 |
| Claude Opus 5 | $5 | 10% | $0.50 |
| Claude Sonnet 5 | $2 | 10% | $0.20 |
| Claude Haiku 4.5 | $1 | 10% | $0.10 |
Fable 5.1 is twice the base input price of Opus 5 and half the cache-read price. For an agent loop, that ordering is the one that matters, because an agent loop is mostly cache reads.
Take a coding agent with a 200K-token cached prefix — repo map, system prompt, tool schemas, accumulated transcript — that runs 40 turns. Cache reads alone are 40 × 200K = 8M tokens:
- Fable 5.1: 8 × $0.25 = $2.00
- Opus 5: 8 × $0.50 = $4.00
That is only the cache-read line. Fresh input and output still bill at full rate, where Fable 5.1 is the more expensive model ($10/$50 vs $5/$25), so the total comparison depends on your ratio of cached prefix to new tokens per turn. The point is that you cannot rank these two models by cost from the headline numbers. You have to know your cache hit profile.
Anthropic’s own framing of the change is roughly 25% savings on typical workloads and up to about 45% on highly agentic ones. Treat those as vendor figures and measure your own; the per-token rates above are the part you can verify directly.
Batch API is still 50% off, and stacks.
Effort defaults differ by surface
Fable 5.1’s default effort is High in Claude Code and Medium on Claude.ai and Cowork. If you are comparing output quality between your terminal and the web app and concluding the model is inconsistent, check this first — you are probably comparing two different effort settings, not two different days.
There is also a new per-message effort beta, which means effort no longer has to be a property of the whole session. For a harness, the useful pattern is low effort for mechanical turns (file reads, test reruns, formatting) and high effort for the turns that actually require planning.
Breaking changes in Fable 5.1 that hit harnesses, not chats
These are called out in the platform overview as breaking relative to Fable 5. They are easy to miss because they mostly do not affect interactive chat use — they affect programmatic orchestration.
Forced tool use now errors. If your harness pins tool_choice to a specific tool to guarantee structured output, that pattern needs revisiting. This is the most likely source of a hard failure on migration.
Earlier models cannot read Fable 5.1’s thinking blocks. If you run a multi-model pipeline — Fable 5.1 plans, Sonnet 5 or Haiku 4.5 executes — you cannot pass the thinking blocks downstream. Whatever the planner needs to communicate has to be in ordinary output content.
Editing earlier turns invalidates thinking. Any harness that rewrites conversation history — transcript compaction, retry-with-correction, replaying a run with one turn patched — loses the thinking state when it does. If you built a replay or rewind feature, test it explicitly against Fable 5.1 rather than assuming it carries over.
On the additive side, the same release adds turn-scoped system messages, a display: "updates" progress mode, the cheaper cache reads described above, and content provenance. Turn-scoped system messages are the one I would reach for first: they let you inject per-step instructions without permanently mutating the cached system prompt, which is exactly the thing that used to force a cache invalidation.
Published benchmarks
From the Fable 5.1 announcement. These are Anthropic’s numbers, cited as their claims — not independent measurements, and not something this site reproduced.
| Eval | Fable 5.1 | Fable 5 | Opus 5 |
|---|---|---|---|
| Terminal-Bench-Science 0.1 | 52.6% | 24.7% | 29.0% |
| Terminal-Bench 4.0 | 55.8% | 42.0% | 52.3% |
| GDPval-AA v2 | 1853 | 1723 | 1824 |
| OSWorld 2.0 (partial / strict) | 77.9% / 41.7% | 72.9% / 36.1% | 75.4% / 39.6% |
| HLE (no tools / with tools) | 60.9% / 65.0% | 57.8% / 63.8% | 56.6% / 63.6% |
| AutomationBench | 31.4% | 17.1% | 26.9% |
| CursorBench 3.2.0 | 73.4% | 70.5% | 70.0% |
The gap that stands out for agent builders is AutomationBench and Terminal-Bench-Science, where the reported margin over Opus 5 is much wider than on the general coding evals. CursorBench, by contrast, shows Fable 5.1 and Opus 5 within a few points. If your workload looks like ordinary editor-scoped coding, the published numbers do not obviously justify the higher base rate; if it looks like long-horizon terminal automation, they do.
Platform behavior your harness should expect
Automatic safety fallbacks. A beta introduced with Opus 5 adds automatic fallbacks when safety classifiers flag Opus 5 or Fable 5. Biology-related Fable blocks now route to Opus 5 rather than Opus 4.8. In practice this means the model that answers may not be the model you requested, so log the responding model rather than assuming your configured one.
Cyber classifier scope. Anthropic’s published policy allows finding vulnerabilities in source code, and blocks binary-based scanning, penetration testing, and exploit generation. If you are building security tooling, that line determines what you can automate. I am describing the policy boundary, not how to work around it.
Mid-conversation tool changes. Also introduced with Opus 5: changing tool definitions mid-conversation without invalidating the prompt cache. For an agent that swaps tool sets between phases, this removes a real cost penalty that previously argued for a single bloated tool list.
Text watermarking
Since August 14, 2026, future Claude models watermark text output, for EU AI Act and Code of Practice compliance. It is global at launch, with a transition period for models released before 2026-08-02.
The engineering-relevant properties, per the announcement: no practical quality impact, no hidden characters, no extra tokens, and not traceable to a person, organization, or conversation. A detection API is coming.
Three practical implications. Your token budgets and cost models are unaffected — there is no token overhead to account for. Your text-processing pipeline does not need to sanitize anything, because there are no hidden characters to strip. And the watermark is not an audit trail: it does not let you attribute a given output back to a specific user or session, so it is not a substitute for your own logging.
Surface changes worth knowing
Claude in Chrome is generally available on every paid plan, with a safety classifier on autonomous browser actions and admin domain allowlists for enterprise.
Memory now works the same across chat and Cowork, with topic-level edit and delete. Sensitive topics are off by default. Memory is on by default for Free, Pro, and Max; Team and Enterprise are admin-gated and off until a user enables it. If you are writing internal guidance for a team account, that last clause is the one people get wrong.
Sources
- Models overview — https://docs.anthropic.com/en/docs/about-claude/models
- Pricing — https://platform.claude.com/docs/en/about-claude/pricing
- Choosing a model — https://platform.claude.com/docs/en/about-claude/models/choosing-a-model
- Claude Fable 5.1 and Mythos 5.1 — https://www.anthropic.com/claude-fable-and-mythos-5-1
- Fable 5.1 platform overview — https://platform.claude.com/docs/en/models/fable-5-1/overview
- Claude Opus 5 — https://www.anthropic.com/news/claude-opus-5
- Opus 5 platform overview — https://platform.claude.com/docs/en/models/opus-5/overview
- Claude text watermarking — https://www.anthropic.com/news/claude-text-watermark
- Claude in Chrome GA — https://claude.com/blog/claude-in-chrome-generally-available
- Memory everywhere — https://claude.com/blog/claudes-memory-works-everywhere-and-you-decide-whats-in-it
- Claude Code docs — https://code.claude.com/docs/en/overview
Model IDs, prices, and dates change. Check the pricing and models pages before you rely on any figure here.