Claude Opus 5.5 vs Fable 5.1 vs Sonnet 5: Which Should You Use?

The practical Claude choice is no longer a three-way contest between Fable 5, Opus 4.8, and Mythos 5. For normal app, Claude Code, and API buyers, the current decision is Sonnet 5 versus Opus 5.5 versus Fable 5.1. Opus 5 matters as an upgrade predecessor, while Mythos 5.1 remains an access-gated security variant. The practical choice is between Sonnet 5 for fast daily execution, Opus 5.5 for high-intelligence coding and analysis, and Fable 5.1 for specialized long-horizon autonomy.

The short answer is to start with Sonnet 5 when responsiveness and volume dominate, move to Opus 5.5 when difficult work makes failed attempts or extra human review expensive, and reserve Fable 5.1 for workloads where a matched evaluation proves that its additional capability is worth the higher price and stricter operating boundary. Anthropic's own model guide positions Opus 5.5 as its frontier flagship for coding and agentic work.

Start with the job, not the model name

Sonnet 5 is the throughput choice. Anthropic positions it as the best combination of speed and intelligence, lists it as fast relative to the other current models, and prices it at $3/$15 per million tokens. It is the sensible baseline for interactive assistants, routine coding, high-volume extraction, and agent loops where latency or token spend compounds across many calls.

Opus 5.5 is the quality-first daily model. Anthropic positions it as its flagship model, offering Always-on Adaptive Thinking, "Less Chatter" conciseness, and 66.4% on Terminal-Bench 4.0. At $4/$20 per million tokens, it drops API pricing by 40% compared to Opus 5's total completion costs, making it the strongest starting point for complex agentic coding and professional analysis.

Fable 5.1 is the capability-ceiling route. Anthropic describes it as targeting long-running autonomous agents, but it costs $10/$50 per million tokens—2.5 times Opus 5.5's base input and output rates. It also carries model-specific retention requirements and stricter safeguards. That makes Fable a specialized escalation, not the automatic premium default.

Anthropic publishes benchmark and safety evaluations for all three models. Those results are useful vendor evidence about intended positioning, but they are not independent comparative testing. Use them to form an evaluation plan, not to declare a universal benchmark winner.

Specs that change the decision

The shared context and output limits remove a common but misleading selection shortcut: choosing the more expensive model does not buy a larger advertised window among these three. The real differences are capability profile, latency orientation, token price, effort behavior, access, and governance.

Per-token price is also not the same as cost per completed job. A slower or more expensive model can be economical if it needs fewer retries, fewer tool calls, or less review. A cheaper model can become costly if a long agent loop repeatedly fails. Measure input tokens, output tokens, elapsed time, retries, tool calls, and human correction together.

Sonnet 5 has standard US$3 input / US$15 output pricing per million tokens. Opus 5.5 is priced at US$4 input / US$20 output with US$0.40 cache reads, delivering roughly 40% net savings over Opus 5 due to more concise outputs. Fable 5.1 retains US$10/US$50 base rates with US$0.25 cache reads. Because Opus 5.5 is priced closely to Sonnet 5, many teams can justify Opus 5.5 as their primary model.

Choose Sonnet 5 for speed and scale

Choose Sonnet 5 when most tasks are repeatable, latency-sensitive, or high-volume. It is the first model to try for customer-facing chat, broad internal assistants, routine coding changes, extraction pipelines, and agents whose economics depend on many inexpensive steps. Its lower standard price leaves more room for retries, caching, or parallel attempts.

Sonnet 5 is also the easiest family baseline for evaluating whether more capability changes the outcome. Run it at the same effort level and with the same tools as Opus 5.5. If it meets the quality threshold, the faster latency orientation and lower price make the selection straightforward.

Do not promote Sonnet solely because a vendor benchmark approaches a premium model. Match the task, tools and effort, then compare accepted results under the current standard US$2/US$10 rate. There is no longer a separate scheduled higher Sonnet rate to budget for.

Choose Opus 5.5 for difficult everyday work

Choose Opus 5.5 when the task is difficult enough that judgment, persistence, or self-correction matters more than the lowest token rate. Complex repository work, multi-step research, professional document analysis, and enterprise agents with costly failure modes benefit from Opus 5.5's adaptive reasoning and 66.4% Terminal-Bench 4.0 accuracy.

Opus 5.5 is the sweet spot: only $1.00/1M more than Sonnet 5 on input, less than half Fable 5.1's price, and free from Fable-specific retention constraints. Anthropic positions Opus 5.5 as its recommended default for agentic coding in Claude Code and enterprise API deployments.

For latency-sensitive premium work, Anthropic offers Fast mode at about 2.5 times the default speed. That route costs twice the base API price ($8/$40 on Opus 5.5) and can consume usage credits in Claude Code, so compare it with Sonnet 5 rather than treating speed as a free toggle.

Choose Fable 5.1 only for measured headroom

Choose Fable 5.1 when long-horizon autonomy, unusually difficult reasoning, or frontier-level capability is the primary constraint and a representative evaluation shows a material gain over Opus 5.5. The evaluation should use the same prompt, tools, effort, stopping conditions, and review rubric for both models.

Fable 5.1's standard US$10 input / US$50 output rates are 2.5 times Opus 5.5's US$4/US$20, but its US$0.25 cache-read rate is slightly below Opus's US$0.40. Cache writes, uncached input, output, retries and review work all affect total cost. Prefer Fable only when complete task economics and quality assessment support it.

Fable normally requires 30-day retention. Eligible customers explicitly authorized by Anthropic can use transitional ZDR while Enterprise Frontier Safeguards rolls out; confirm the agreement rather than assuming every account qualifies. EFS is planned in phases beginning later in the fall; it is not already universally deployed. Privacy settings for consumer chat are not an API ZDR agreement.

Stricter safety classifiers can also refuse or reroute some requests. In Claude surfaces, Anthropic may fall back from a flagged Fable request to another model; API users need to handle refusal details or configure an eligible fallback path. For cybersecurity, biology, or other sensitive workflows, confirm which model actually answers and whether the resulting behavior remains acceptable.

Effort changes the model you are buying

Fable 5.1, Opus 5.5, and Sonnet 5 all support the low, medium, high, xhigh, and max effort levels. High is the API default. Effort affects thinking, visible output, and tool calls, so it changes both capability and cost; it is a behavioral signal rather than a strict token budget.

Start a family comparison at high effort, then sweep downward or upward only where the workload justifies it. Sonnet at high may beat Opus at low for one task, while Opus at high may avoid retries that make Sonnet more expensive in practice. A model comparison without matched effort is partly an effort comparison.

Opus 5.5 features native Always-on Adaptive Thinking, eliminating rigid thinking token budgets. Sonnet 5 and Fable 5.1 also use adaptive thinking by default. Applications should tune the effort parameter rather than attempting to disable internal reasoning on complex tasks.

Max effort is not a default buying recommendation. It allows unconstrained token spending for the deepest work. Use it only when a measured quality gain matters more than the added latency and cost, and set output limits with enough room for thinking and tool use.

Claude app, Claude Code, API, and cloud routes

Model access and billing depend on the surface and account. Fable 5 and 5.1 share up to 50% of regular weekly usage on Max and Premium seats; Pro and Standard seats use paid usage credits from the start. This is not an extra allowance for each model. The earlier Fable 5 promotion did not include Fable 5.1 and is not renewed by this release. Direct API use and usage-based Enterprise follow their own metered rates.

Claude Code's model selector supports Sonnet 5, Opus 5.5, Fable 5.1, and Opus 5. Use the /model menu or a model flag to make the choice explicit, then check /status so an evaluation does not accidentally run against a different model or billing route. Fable 5.1 requires Claude Code 2.1.255 or later; the fable alias points to it unless overridden, but neither Fable model is an account-type default.

A paid Claude app plan does not include direct Claude API access. Console and API usage are billed separately, and Claude Code can use either a supported subscription allocation or API credits depending on authentication and account state. Keep app entitlement, Claude Code usage, and direct API spend as separate budget lines.

For developers, all three current models are available through the Claude API, Claude in Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Model IDs, feature support, fallback behavior, data processing, and regional availability can differ by route. Anthropic is the processor for its native API, Claude Platform on AWS, and Claude in Microsoft Foundry; cloud-provider terms govern Amazon Bedrock and Google Cloud processing.

Opus 5 and Mythos 5.1 are boundaries, not peer choices

Keep Opus 5 as a predecessor baseline for existing integrations rather than the current default comparison. Fable 5.1 needs its own API migration check: use claude-fable-5-1 and remove forced tool choices. Older models cannot read its thinking blocks; on a model switch the API drops incompatible blocks and the older model re-plans.

Mythos 5.1 shares Fable 5.1's specifications and pricing but is not a general buyer option. Anthropic limits it to approved Project Glasswing customers for defensive cybersecurity workflows, with no self-serve signup. Unless an organization already has that access path, compare Sonnet 5, Opus 5.5, and Fable 5.1 instead.

Run a workload evaluation before switching

Build a small evaluation set from work that is valuable, difficult, and representative. Include routine tasks that expose needless premium spend, difficult tasks that expose quality differences, and sensitive tasks that can trigger safeguards or governance constraints.

For each candidate, keep prompts, tools, context, effort and stopping rules matched. Record accepted output, human corrections, retries, tool calls, token mix, elapsed time and fallback events. Use current standard Sonnet pricing and each model’s distinct cache-read/write rates; do not revive the cancelled Sonnet price increase.

Choose the least expensive route that reliably clears the quality, latency, and governance thresholds. Start with Sonnet 5 for throughput work, Opus 5.5 for difficult daily work, and Fable 5.1 only where its measured headroom pays for the higher price and operating constraints. Re-run the evaluation when prompts, tools, effort settings, or cloud routes change.