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 was correct at launch, but the pricing landscape shifted significantly on July 24 when Opus 5 launched at $5/$25 per million tokens with near-Fable-5 intelligence. The relevant comparison is now Fable 5 vs Opus 5, not Fable 5 vs Opus 4.8.

The useful question remains: what does each model cost per merged PR, per shipped feature, per closed ticket? Those numbers look different across Fable 5, Opus 5, Sonnet 5, and Opus 4.8 - and in some workflows the premium for Fable 5 still pays for itself.

Last updated: July 28, 2026

What Changed on July 28, 2026#

• Opus 5 launched at $5/$25 per million tokens - exactly Opus 4.8 pricing but with near-Fable-5 capability. This resets the entire cost-per-task comparison. Teams that balked at Fable 5's 2x premium now have Opus 5 as a middle option at zero price premium over Opus 4.8.
• Fable 5 went API-only on July 9, 2026. Subscription access via Claude.ai Pro/Max tiers was withdrawn. Fable 5 is now available only through the Anthropic API and remaining Max plan access (limited).
• Sonnet 5 launched with promotional pricing ($2/$10 per million tokens through Aug 31), replacing Sonnet 4.6 as the budget-friendly Claude option.
• Fable 5 tokenizer note confirmed: the 30% higher token count vs pre-4.7 models remains a real factor in per-task cost calculations.

The Numbers Up Front#

Quick answer: Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens on the Claude API - double Opus 4.8's $5/$25 and double Opus 5's $5/$25. Since Opus 5 launched at the same price as Opus 4.8 but with near-Fable-5 intelligence, the value argument for paying Fable 5's premium is now narrower. Per Anthropic's models overview (accessed July 28, 2026).

One caveat the rate card hides: Fable 5 uses the tokenizer introduced with Opus 4.7, which produces roughly 30% more tokens for the same text than pre-4.7 models, so teams re-baselining from older models pay the higher rate on a larger token count, not just a clean 2x multiplier. Opus 5 and Sonnet 5 use the same tokenizer, so token counts are comparable across the current lineup.

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

Fable 5 costs 2x Opus 5 and 5x Sonnet 5 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. Anthropic's pricing page (accessed July 28, 2026) confirms the current lineup. GPT-5.6 Sol pricing at $5/$30 per million tokens provides a comparable alternative for teams evaluating both providers.

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 5: approximately $0.003 per review (promo pricing)
• Opus 4.8: approximately $0.025 per review
• Opus 5: approximately $0.025 per review (same rate as Opus 4.8)
• Fable 5: approximately $0.10 per review

At $0.10 per review, Fable 5 is objectively more expensive than Sonnet. But Opus 5 at $0.025 offers near-Fable-5 review quality at the same price as Opus 4.8, making it the sweet spot for code review. If your team does 40 reviews a day, that is $4 for Fable 5, $1 for Opus 5, or $0.12 for Sonnet 5. The cost difference is real across all three.

Full feature build: a Next.js pricing page

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

• Sonnet 5 (promo): ~$0.09 in tokens, plus roughly two hours of cleanup
• Opus 4.8: ~$0.94 in tokens, with moderate cleanup
• Opus 5: ~$0.94 in tokens, often near-production-ready (matches Fable 5 quality tier)
• Fable 5: ~$4.18 in tokens, often production-ready

Opus 5 changes the math dramatically. At Opus 4.8 pricing with near-Fable-5 quality, the premium over Sonnet is $0.85 versus $3.24 for Fable 5. If your time is worth anything above $30 an hour, the Opus 5 premium pays for itself if it saves 1-2 minutes of cleanup - a much lower bar than Fable 5's requirement.

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. With Opus 5 at the same $5.40 rate, you get approximately 90% of Fable 5's capability at 23% of the cost. This makes Opus 5 the default choice for most async agent work, reserving Fable 5 for edge cases where only the absolute best output quality suffices.

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. But Opus 5 now handles many of these cases at half the token cost, narrowing Fable 5's exclusivity.

Retries and rework compound. If Opus 5 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. Opus 5's pass rate is close enough to Fable 5 on most coding tasks that this advantage is smaller than it was against Opus 4.8.

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. Opus 5 can handle similar tasks at half the API cost, making it the practical choice for most long-running agents.

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 $2 per million input tokens (Fable 5 vs Sonnet 5 promo) is material - Fable 5 costs 5x more per task. Sonnet 5 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 5 produces equally usable results at one-fifth the cost of Fable 5 (promo pricing).

High-volume pipelines. Teams running thousands of API calls per day need to model cost per call, not cost per task. At that volume, a 5x difference between Fable 5 and Sonnet 5 is a major line item. The right architecture often uses Haiku or Sonnet 5 for the high-volume path and routes hard edge cases to Opus 5 or Fable 5.

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 5'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.

Fable 5 Access Model: API-Only Since July 9#

Fable 5 launched via API on June 9, 2026, with subscription access phased in starting June 22. However, on July 9, 2026, Anthropic withdrew Fable 5 subscription access from Claude.ai Pro and Max tiers. Fable 5 is now API-only.

This matters for cost planning because the access model determines your pricing surface:

API direct: $10/$50 per million tokens. Full Fable 5 access. You pay for every token. This is the only way to use Fable 5 as of July 28, 2026. Teams running production pipelines are here.

Max subscription (Claude Code): The $100 and $200 per month Max plans include Opus 5 and Sonnet 5 access without per-token billing up to usage caps. Fable 5 on Max plans is limited - most usage will default to Opus 5, which delivers near-Fable-5 quality at no extra subscription cost. The Claude Code usage limits playbook covers how Max tiers allocate capacity in practice.

Pro subscription ($20/month): Sonnet 5 access with token budgets. Opus 5 available on Max tier upgrades. No Fable 5 access via subscription.

The practical implication: if you want Fable 5 in July 2026, you pay API rates. For most workflows, Opus 5 gives you approximately 90% of Fable 5's capability at the same per-token price as Opus 4.8 ($5/$25), making it the more practical choice for all but the most demanding tasks.

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.

At Opus 5 rates (same tasks):

• 5 large builds at $0.94 each = $4.70
• 20 code reviews at $0.025 each = $0.50
• 2 agent sessions at $5.40 each = $10.80

Total: ~$16/day, or roughly $340/month. Opus 5 delivers approximately 90% of Fable 5's quality at 26% of the cost.

Step 3: Apply routing

Most cost-optimized setups use Opus 5 for tasks where quality matters, Fable 5 for edge cases where only the absolute best output suffices, and Sonnet 5 or Haiku for volume tasks. The blend typically lands heavy users in the $200 to $500 per month range at API rates (with Fable 5 on critical paths), or $100 to $350 per month with Opus 5 as the primary model. Max subscription plans at $200/month absorb the upper end for individual developers using Opus 5 as their primary model.

FAQ#

How much does Claude Fable 5 cost?#

Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens on the Claude API, confirmed in Anthropic's pricing page (accessed July 28, 2026). That is 2x Claude Opus 5 ($5/$25) and 5x Claude Sonnet 5 promotional pricing ($2/$10, through Aug 31). Budget for the tokenizer too: Fable 5, Opus 5, and Sonnet 5 all use the tokenizer introduced with Opus 4.7, which produces roughly 30% more tokens for the same text than pre-4.7 models.

Can I still use Fable 5 on a Claude Pro or Max subscription?#

Not in a meaningful way. On July 9, 2026, Anthropic withdrew Fable 5 subscription access from Claude.ai Pro and Max tiers, so Fable 5 is now effectively API-only at $10/$50 per million tokens. Pro has no Fable 5 access, and Fable 5 on Max plans is limited, with most usage defaulting to Opus 5.

Is Fable 5 or Opus 5 the better value for coding?#

For most teams, Opus 5. It launched at $5/$25 per million tokens - the same as Opus 4.8 - with near-Fable-5 capability. In the walkthrough above, the same solo-developer workload costs roughly $1,300 per month at Fable 5 rates versus roughly $340 per month on Opus 5. Fable 5 still earns its premium on long-horizon agentic tasks and high-stakes work where a wrong answer creates significant downstream cleanup.

How do I estimate my monthly Fable 5 bill?#

Work out three things: your task mix, the average tokens per task, and how many tasks run per day. Then apply routing - Opus 5 where quality matters, Fable 5 for edge cases, and Sonnet 5 or Haiku for volume tasks. Measure cost per merged PR rather than cost per million tokens, and set spending caps before running overnight agents.

Official Sources#

All links verified July 28, 2026.

Bottom Line: When to Use Fable 5, When to Use Opus 5, When to Stay on Sonnet#

Use Fable 5 when:

• Your workflow involves long-horizon agentic tasks where only the absolute best output quality suffices
• You are working on high-stakes tasks where a wrong answer creates significant downstream work
• The 5x premium over Sonnet 5 is negligible relative to the cost of failure
• You are willing to pay API rates (Fable 5 is no longer available on subscription plans)

Use Opus 5 when (recommended default for most teams):

• You want near-Fable-5 quality at Opus 4.8 pricing ($5/$25)
• Your tasks involve multi-file refactors, code review, or complex reasoning
• You use Claude Code Max subscription ($100-200/mo) and Opus 5 is included
• You need the strongest price-performance ratio in the current Claude lineup

Use Sonnet 5 (promo pricing through Aug 31) when:

• Your primary workload is interactive chat, scaffolding, or volume classification
• You are running thousands of API calls daily and the 5x cost difference vs Fable 5 is a material budget line
• Your tasks are well-scoped and Sonnet 5 produces acceptable results consistently
• You are in a cost-optimization phase

The decision framework changed with Opus 5's launch. Opus 5 now handles most of what teams were paying Fable 5 for, at half the token cost. Reserve Fable 5 for the small fraction of tasks where the marginal quality gain justifies the premium.

For teams planning their budget, Anthropic's own recommendation is to run Opus 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. See the frontier model API pricing tracker for the full multi-provider comparison.

Continue Reading#

• Beyond the Pelican Test: Opus 5 Renders the Lord of the Rings With a 1M-Token Budget