Claude Fable 5 Pricing: Real Cost Per Task vs Opus 4.8, GPT-5.5 and Codex

Fable 5 launched on June 9, 2026 at $10 per million input tokens and $50 per million output tokens. The immediate reaction in developer circles was "that is 2x Opus 4.8." That framing is correct on the rate card and mostly irrelevant to whether the upgrade makes economic sense for your workflow.

The useful question is not what Fable 5 costs per token. It is what Fable 5 costs per merged PR, per shipped feature, per closed ticket. Those numbers look different, and in some workflows they invert the premium entirely.

Last updated: June 10, 2026

The Numbers Up Front

Here is how Fable 5 sits against the models most teams are actually choosing between right now.

Fable 5 costs exactly 2x Opus 4.8 at list price. Anthropic's announcement notes this is less than half the cost of the Claude Mythos Preview that preceded it, so the trajectory for the top tier is downward. GPT-5.5 pricing has varied by access tier; Codex CLI users on ChatGPT Pro have a separate compute allocation that does not map cleanly to per-token rates, making direct comparison harder.

For a direct breakdown of how these numbers stack up against Cursor, Codex, and other tools across the full spectrum, the AI coding tools pricing comparison has the side-by-side view.

Why List Price Lies: Cost-Per-Task Is What Matters

Token rates tell you the cost per unit of compute. They do not tell you the cost per unit of work. Those two numbers diverge whenever a more capable model does the same job in fewer turns, with less cleanup, or with zero retries.

Three common coding tasks illustrate the gap.

PR review on a 200-line change

A shallow review costs roughly the same across models because the output is short. A thorough review that flags the subtle auth bug in line 47 may only come from the more capable model, making the cheaper model's savings illusory if you are reviewing the diff again anyway.

• Sonnet 4.6: approximately $0.01 per review
• Opus 4.8: approximately $0.025 per review
• Fable 5: approximately $0.10 per review

At $0.10 per review, Fable 5 is objectively more expensive than Sonnet. If your team does 40 reviews a day, that is $4 versus $0.50. The cost difference is real. The question is whether the review quality reduces post-merge defects enough to justify it.

Full feature build: a Next.js pricing page

This is where the cost-per-outcome math starts to shift. Research from AY Automate found the following when building the same feature from scratch:

• Sonnet 4.6: ~$0.25 in tokens, plus roughly two hours of cleanup
• Opus 4.8: ~$0.94 in tokens, with moderate cleanup
• Fable 5: ~$4.18 in tokens, often production-ready

If your time is worth anything above $30 an hour, the $3.93 premium on Fable 5 pays for itself if it saves two hours of cleanup. If the Sonnet output is good enough with light editing, the premium does not.

Multi-hour async agent run

Long-running agents expose the largest absolute dollar differences. The same research estimates async agent runs at around $5.40 for Opus 4.8 versus $23.70 for Fable 5 on comparable tasks. This is a $18.30 gap per run. Whether that is worth it depends entirely on whether the Fable 5 output is merge-ready and the Opus 4.8 output requires another iteration.

Anthropic notes that Fable 5 is more token-efficient than past Claude models, which means it sometimes costs fewer output tokens to complete the same task even at a higher rate. This efficiency gain does not show up in rate card comparisons.

The 2x Premium Case: Where Fable 5 Wins on Cost

The premium pays off in workflows where quality failure is expensive.

Retries and rework compound. If Opus 4.8 produces an output that needs one correction loop, the effective cost is two API calls. If Fable 5 produces the right output on the first pass, the math is closer than the 2x rate suggests.

Long-horizon agentic work. Fable 5 is Anthropic's stated choice for tasks that take a person hours, days, or weeks. Stripe reported it compressed months of engineering on a 50-million-line Ruby migration. At that scope, the per-token cost is a rounding error compared to engineering hours.

High-stakes accuracy tasks. Finance, legal, and compliance workflows where a wrong answer creates downstream work are the natural home for the most capable model. Fable 5 achieved the highest score on Hebbia's Finance Benchmark. At $0.01 per review task, the premium over Sonnet is negligible relative to the risk of a missed issue.

Vision-heavy pipelines. Fable 5 inherits Opus 4.7's high-resolution vision support (up to 2,576 pixels on the long edge, versus 1,568 on Opus 4.6). Screenshot analysis, design implementation from mocks, and document understanding tasks benefit from this, and the per-task cost difference is small for single-image workflows.

The 2x Premium Bust: Where Opus 4.8 or Sonnet Is the Right Call

The premium does not pay when the task does not require Fable-level capability.

Classification and extraction. Single-turn tasks with structured outputs are cost-sensitive by nature. If you are classifying 100,000 support tickets, the difference between $10 and $5 per million input tokens is material. Sonnet or Haiku handles most classification reliably, and the output is either correct or not - there is no cleanup multiplier.

Boilerplate and scaffolding. Generating CRUD routes, configuration files, or test stubs does not require the most capable model. The output is deterministic enough that Sonnet produces equally usable results at one-third the cost.

High-volume pipelines. Teams running thousands of API calls per day need to model cost per call, not cost per task. At that volume, the 2x difference is a meaningful line item. The right architecture often uses Haiku or Sonnet for the high-volume path and routes hard edge cases to a capable model.

Interactive chat with fast iteration. Conversational back-and-forth where the user corrects course each turn does not benefit from Fable's long-horizon autonomy. Sonnet's lower latency and cost profile fit better here.

The AI coding tools pricing Q2 2026 post has a broader breakdown of how to route different task types across model tiers.

June 22 Subscription Cliff: What It Actually Costs After

Fable 5 launched via API on June 9, 2026, but subscription access is being phased in. According to Anthropic's rollout plan, Claude.ai Pro users get access starting June 22, 2026.

This matters for cost planning because the three pricing surfaces behave differently:

API direct: $10/$50 per million tokens. Full Fable 5 access from day one. You pay for every token. Teams already running production pipelines are here.

Max subscription (Claude Code): The $100 and $200 per month Max plans include model access without per-token billing up to usage caps. Fable 5 access on these tiers depends on the rollout schedule. The Claude Code usage limits playbook covers how Max tiers allocate capacity in practice.

Pro subscription ($20/month): Access post-June 22. Token budgets apply. Users on Reddit have documented Fable 5 consuming the Max 20x plan at rates that surprised them - the model's capability means tasks that previously took 5 turns take 2, but each turn uses more tokens due to longer, more thorough outputs.

The practical implication: if you are planning to use Fable 5 heavily through a subscription, budget for a tier increase. The $200 Max plan is where uncapped heavy use lands, and it remains a significant value compared to direct API billing for high-volume individual developers.

Uber's $1,500 Per Month Cap Lesson

The clearest signal about what enterprise teams are actually spending came from Uber's internal policy. According to reporting covered by Simon Willison, Uber capped per-employee spending at $1,500 per month per agentic coding tool - covering Cursor, Claude Code, and similar tools independently.

This cap emerged because Uber exhausted its entire 2026 AI budget within four months before controls were in place. At $1,500 per employee per tool, with two tools, the annualized cost is $36,000 per engineer - roughly 11% of median Uber software engineer compensation.

The lesson for planning teams is not that $1,500/month is the right cap. It is that per-token billing on agentic workflows without guardrails produces costs that do not match 2025 budget assumptions. The models got more capable, the agents got more autonomous, and token consumption scaled with both.

Teams adopting Fable 5 for agentic work should model worst-case consumption before setting budgets. A developer running async agents for 4 to 6 hours daily at Fable 5 rates could exceed $500/month in API costs. The 400 dollar overnight bill post walks through how runaway agent costs happen and how to prevent them.

Cost Calculator Walkthrough

Estimating your monthly bill requires knowing three things: task mix, average tokens per task, and how many tasks run per day.

Step 1: Categorize your workflow

Step 2: Estimate daily volume per type

A solo developer doing active feature work might run 5 large feature builds, 20 code reviews, and 2 to 3 longer agent sessions per day.

At Fable 5 rates, that is roughly:

• 5 large builds at $4 each = $20
• 20 code reviews at $0.10 each = $2
• 2 agent sessions at $20 each = $40

Total: ~$62/day, or roughly $1,300/month at full pace. This is consistent with AY Automate's estimate that an async-agent power user lands around $450/month at mixed model routing, and considerably higher at pure Fable 5.

Step 3: Apply routing

Most cost-optimized setups use Fable 5 for tasks where quality determines the outcome, and Sonnet or Haiku for volume tasks where the output is either right or easily corrected. The blend typically lands heavy users in the $200 to $600 per month range at API rates, with Max subscription plans absorbing the upper end for individual developers.

Official Sources

Bottom Line: When to Upgrade, When to Stay on Opus

Upgrade to Fable 5 when:

• Your workflow involves long-horizon agentic tasks where the quality of the output determines whether you need another pass
• You are working on high-stakes tasks where a wrong answer creates significant downstream work
• You already hit quality ceilings on Opus 4.8 for complex reasoning, code, or vision tasks
• You are on a Max subscription plan and Fable 5 access is included in your tier

Stay on Opus 4.8 when:

• Your primary workload is interactive chat, scaffolding, or volume classification
• You are running hundreds or thousands of API calls daily and the 2x cost difference is a meaningful budget line
• Your tasks are well-scoped and Opus 4.8 produces acceptable results consistently
• You are in a cost-optimization phase and the marginal quality gain does not justify the rate increase

The 2x rate premium is real. The cost-per-outcome premium is situation-dependent, and for the right tasks it disappears entirely. The decision is less about which model is better and more about whether the tasks you are running are the ones where the capability gap shows up as billable savings.

For teams planning the transition, Anthropic's own recommendation is to run Fable 5 at effort: "high" or "xhigh" for coding and agentic work, which means higher per-call token consumption than prior models. Budget accordingly, set spending caps before running overnight agents, and measure cost per merged PR rather than cost per million tokens.