Sonnet 5 Ships. Fable 5 Returns. Now What?

Anthropic’s framing is simple: Claude Sonnet 5 is “the most agentic Sonnet model yet.” It can plan, use tools like browsers and terminals, and run autonomously at a level that recently required bigger and more expensive models. It is now the default model on Free and Pro, available across paid plans, live in Claude Code, and available through the API as claude-sonnet-5.

The pricing is simple at first glance. Through August 31, Sonnet 5 costs $2 per million input tokens and $10 per million output tokens. After that, it moves to $3 input and $15 output. Opus 4.8 is $5 input and $25 output. So yes, per token, Sonnet 5 is cheaper.

But per-token pricing is not the same thing as per-task cost.

Anthropic says Sonnet 5 is a substantial improvement over Sonnet 4.6 on reasoning, tool use, coding, and knowledge work, and that its performance is close to Opus 4.8 at lower prices. Their own charts call out medium effort as the efficient tier, while higher effort can match Opus 4.8 on some tasks. That wording matters: “on some tasks” is not “replaces Opus.”

The practical concern is token efficiency. Anthropic says Sonnet 5 uses an updated tokenizer, and the same input can map to roughly 1.0-1.35x more tokens depending on the content. The introductory price is meant to make that transition roughly cost-neutral. That is a very different claim from “this is simply cheaper.”

If the model needs more tokens to do the same job, and if higher effort tiers burn enough budget to chase Opus-like performance, the comparison changes. Sonnet 5 medium may be a useful workhorse. Sonnet 5 max or xhigh may be a worse deal than Opus 4.8 for harder tasks. That was the recurring complaint in the developer community too: cheaper per token, but sometimes more expensive per completed task.

This is where the old “just use the cheaper model” instinct breaks. For real workflows, the unit of cost is not a token. It is a resolved ticket, a working migration, a correct review, or a task that did not need three retries and a human rescue.

Early hands-on testing is pretty positive on design and coding. Sonnet 5 looks strong for one-shotting applications, cleaner web design, single-file interactive demos, and visual UI work. Some developers also said the coding is better than 4.6, and several liked the idea of a less verbose day-to-day alternative to Opus.

That is the case for Sonnet 5: a capable execution model for everyday agentic work, especially when you do not want to spend Opus credits on tasks that do not need Opus judgment. The negative signal is not that Sonnet 5 is weak. It is that the experience is uneven: more refusals, more second-guessing, more moralizing, more over-explaining, and more friction around prompts older models handled normally.

That is the tradeoff across these releases. The model gets safer, more careful, and more constrained. Sometimes that is good. Sometimes it is friction with better branding.

Anthropic says it did not deliberately train Sonnet 5 on cybersecurity tasks. Sonnet 5 can do routine, non-harmful cyber work, but it performs substantially worse than Opus 4.8 and Mythos 5 on dangerous cyber evaluations like exploit development. It also launched with cyber safeguards enabled by default, though less strict than the Fable 5 safeguards.

From a policy seat, this makes sense. From an engineering seat, it is complicated.

If a model is intentionally weaker at exploit development, fine. The risk is real. But if the safety layer confuses legitimate security review with dangerous cyber work, the model becomes less useful precisely where secure software work needs help.

That is the question developers immediately asked: will it refuse security audits the way Fable did? For teams building real software, that question matters more than most of the launch benchmark table.

Fable 5 coming back changes the read of Sonnet 5. Without Fable, Sonnet 5 is the new Claude release. With Fable back in the room, Sonnet 5 becomes the cheaper model you route to when you do not need the frontier one.

That may be the product ladder Anthropic wanted all along: Sonnet for everyday execution, Opus for higher judgment, Fable for the work that actually needs frontier capability. But after the original Fable pullback, “Fable is back” raises a different question: back for whom, under what controls, at what price, and with what guarantees that it will still be there next week?

The facts are cleaner now than they were last week. Anthropic says the June 12 suspension happened because the US government issued an export-control directive requiring Fable 5 and Mythos 5 access to be suspended for foreign nationals, including foreign-national Anthropic employees. With no reliable real-time nationality check, Anthropic disabled both models for everyone.

Anthropic also says the concern was a narrow jailbreak, not a universal one: effectively asking the model to read a specific codebase and fix software flaws. Anthropic argued that this capability was already available from other models, including GPT-5.5, and did not justify recalling a commercial model at that scale.

On July 1, Fable 5 came back globally across Claude Platform, Claude.ai, Claude Code, and Claude Cowork. Mythos 5 did not; access was restored only for approved US organizations while Anthropic works through Glasswing access. For Pro, Max, Team, and select Enterprise users, Fable 5 is included for up to 50% of weekly usage limits through July 7. After that, it moves to usage credits.

The most important change is not model quality. It is the fallback behavior.

Anthropic says it updated the cybersecurity safeguards after conversations with the US government. They say most coding work is unaffected, but also that the new safeguards will flag more harmless requests in the near term. When a request is flagged, users are notified and get a response from Opus 4.8 instead. Biology and chemistry classifiers are also still broader than Anthropic wants.

That is the whole tension. The model is back, but the boundaries around the model are now part of the experience. The visible user experience is a fallback notice: Fable pauses, the request is flagged, and the work continues in Opus 4.8. If a frontier model routes away from itself whenever the prompt gets close to security, biology, chemistry, or anything the classifier associates with those areas, then the classifier is the experience.

We saw the same pattern internally. In an authorized defensive review of a client codebase, a vulnerability-finding prompt was flagged before Fable produced the analysis, and the session continued in Opus 4.8 instead. That is exactly the work where Fable should be valuable for a software team.

The logic is understandable. If Mythos-class capability can accelerate vulnerability discovery, misuse is not hypothetical. But the current tradeoff is over-broad in practice: it reduces risk by blocking the same category of work that makes the model useful for defensive engineering.

That is the uncomfortable part. Fable 5 may be an extremely good LLM, especially for software development and vulnerability analysis. But if legitimate security work falls through to Opus 4.8, then Fable is not actually available for one of its highest-value use cases.

The second issue is usage burn. Developer reaction is split between “we’re so back” and “this is unusable.” Some users say Fable is night-and-day better when it works. Others say it burns through session limits so quickly that it feels more like a demo than a tool, especially when Claude Code spins up multiple Fable subagents.

Some of this is workflow design. Point a frontier model at a repo and let it spawn a swarm of Fable agents, and it will burn budget. But a model can be technically available and practically unusable if the first serious task eats the session.

The 50% weekly-usage cap through July 7 also changes behavior. It makes Fable feel scarce before it even becomes usage-credit gated.

This is where the title’s question mark belongs. Sonnet shipped. Fable returned. Now what?

Maybe. But “great” cannot just mean “the model is powerful.” We already knew that. “Great” has to mean the access model, safeguards, pricing, and reliability are usable enough to build around. On that test, the answer is still unsettled.

The lesson is not “use Sonnet 5” or “avoid Sonnet 5.” The lesson is that model selection is becoming operational work.

Based on the information we have so far, the first thing to do is compare Sonnet 5 Medium directly against Sonnet 4.6 on our agentic workloads. Anthropic’s own reporting, and the analysis around it, point to Sonnet 5 Medium as the most sensible cost-effective cap. Above that, the better move may simply be Opus 4.8.

So test it deliberately. Compare how Sonnet 5 handles your real workflow, find its quirks, and see whether it improves or degrades the way you personally work.

For day-to-day engineering, the current routing shape is probably:

• Use Opus for planning, architecture, hard reasoning, and final judgment by default.
• Test Sonnet 5 Medium for implementation, iteration, UI generation, and bounded agentic tasks.
• Move to Opus 4.8 rather than pushing Sonnet 5 into higher-effort tiers by default.
• Use Fable 5 during the included subscription window for the heaviest work where its capability actually matters.
• Treat defensive security work as a specific test case, because Fable may fall back exactly where we want it most.
• Measure cost per completed task, not cost per token.

For Fable 5 specifically, the short-term play is simple: while it is still included in our subscription plans, use it on the hardest useful work we have. Once it moves behind usage credits, the decision changes. Right now, the window is limited, so use it deliberately.

That last point is the same lesson as issue nine, from the opposite direction. Fable 5 showed the risk of a model being too powerful and too politically sensitive to remain freely available. Sonnet 5 shows the risk of a model being positioned as cheaper and safer while the real workflow cost becomes harder to see.

Both point to the same operating principle: do not build around a model. Build around a routing layer.

The model should be replaceable. The workflow should survive.

Sonnet 5 may become a useful default for a lot of our work. But it has to earn that role in our own tests, especially where security review, instruction following, and cost discipline matter.

And Fable 5 coming back does not erase the lesson from Fable 5 disappearing. If anything, it makes the lesson more urgent. The frontier model may be available again. That does not mean it is stable enough to be the center of the system.

Sources → Anthropic Sonnet 5  ·  Anthropic Fable redeployment  ·  Anthropic suspension statement  ·  Early hands-on testing and developer community discussion