Claude Fable 5.2 Pricing: What the 20% Cut Means per Task vs Fable 5.1, Opus 5.5 and GPT-6 Astra

Research scope: This study compares official list prices on the Anthropic, OpenAI and Google pricing pages, read on November 10, 2026, one day after Claude Fable 5.2 was released. Workload costs are calculated from those list prices and a stated token mix (a short chat turn, one 400K-token prompt and a cache-heavy coding agent session); they are not measured invoices. Benchmark scores are vendor-reported by Anthropic.

Short answer: Claude Fable 5.2 costs $8 for input and $40 for output per 1M tokens, exactly 20% below Claude Fable 5.1 rates of $10 and $50. Cache reads run at $0.20 per 1M tokens, creating the Cache Gap that makes a long coding agent session 35.9% cheaper than GPT-6 Astra. Even with that discount, Fable 5.2 still costs 1.85 times as much as Claude Opus 5.5 for typical agent workloads.

What did Anthropic change on November 9?

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026. Fable 5.2 uses the API identifier claude-fable-5-2 and is available to all developers on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Mythos 5.2 carries the same list prices as Fable 5.2, but Anthropic restricts its access to organizations verified through specific programs such as the Cyber Verification Program and the Life Sciences Verification Program.

Both models maintain a 1M-token context window, support a maximum output of 128K tokens, and have a knowledge cutoff of June 2026. Anthropic states that both versions generate output up to 30% faster than the 5.1 generation. Developers using Fable 5.1 do not face an immediate cutoff, as Anthropic will keep Fable 5.1 available as a legacy model until at least September 1, 2027.

How much does Claude Fable 5.2 cost?

Claude Fable 5.2 standard tier pricing sits at $8 per 1M input tokens and $40 per 1M output tokens, compared to $10 and $50 for Fable 5.1. Anthropic set cache read pricing at $0.20 per 1M tokens, down from $0.25 on Fable 5.1. Cache write pricing dropped in equal measure, moving to $10 per 1M tokens for 5-minute writes and $16 per 1M tokens for 1-hour writes, compared to $12.50 and $20 on the prior release.

The Batch API offers half-price execution at $4 per 1M input tokens and $20 per 1M output tokens. For interactive applications that need lower latency, fast mode costs $16 per 1M input tokens and $80 per 1M output tokens. Anthropic states that fast mode speeds output generation up to 2.5 times, but developers cannot combine fast mode with the Batch API discount.

Every listed price point for Fable 5.2 sits exactly 20% below the equivalent tier on Fable 5.1. Anthropic does not charge a long-context surcharge, meaning requests that use the entire 1M-token window bill at the base $8 and $40 rates. For developers using external gateway routing, Tokely lists Fable 5.2 at $6 input and $30 output per 1M tokens.

How does that compare with GPT-6 Astra, Opus 5.5 and the rest?

The frontier tier splits into two clear price brackets across the primary model providers. Claude Fable 5.2 lists at $8 input, $40 output, and $0.20 for cache reads per 1M tokens, while GPT-6 Astra lists at $10 input, $50 output, and $1.00 for cache reads. Claude Fable 5.1 remains listed at $10 input, $50 output, and $0.25 cache reads, matching Astra on raw input and output base rates.

Beneath the top tier, Claude Opus 5.5 lists at $4 input, $20 output, and $0.20 cache reads per 1M tokens. The mid-tier options run lower still: Claude Sonnet 5.5 lists at $2 input, $10 output, and $0.10 cache reads, matching GPT-6.1 Sol at $2 input, $10 output, and $0.10 cache reads. Gemini 4 Argon currently lists at an introductory price of $2 input, $10 output, and $0.10 cache reads, but Google stated this price will double to $4 input, $20 output, and $0.20 cache reads on December 1, 2026.

These list prices make Fable 5.2 the cheapest of the two top-priced flagships by 20% on input and output. At the same time, Fable 5.2 remains exactly twice the list price of Opus 5.5 for uncached tokens. How those differences translate to production invoices depends directly on how each platform bills cached context.

Standard-tier list prices per 1M tokens on official pricing pages, read November 10, 2026. Gemini 4 Argon is shown at its introductory price, which doubles on December 1, 2026.

The Cache Gap

The pricing difference between flagships expands significantly when evaluating prompt caching mechanisms. We define "the Cache Gap" as follows: Claude Fable 5.2 and GPT-6 Astra differ by 20% on list input and output prices ($8 vs $10 input, $40 vs $50 output per 1M tokens), but Fable 5.2 reads cached tokens at $0.20 per 1M against $1.00 for Astra, five times less. In a cache-heavy coding agent session the bill differs by 35.9% ($17.44 vs $27.20), not 20%, because cache reads are 26.5% of the Astra bill and 8.3% of the Fable 5.2 bill.

Cache read rates vary widely across model catalogs. GPT-6 Astra charges $1.00 per 1M tokens, Fable 5.1 charges $0.25, Fable 5.2 charges $0.20, and Opus 5.5 charges $0.20. Lower down the stack, Sonnet 5.5, GPT-6.1 Sol, and Gemini 4 Argon all bill cache reads at $0.10 per 1M tokens.

Anthropic bills cache reads on Fable models at 0.025 times base input, while OpenAI bills Astra cache reads at 0.1 times base input. Because agent tools resend the entire accumulated context on every turn, cached input tokens account for the vast majority of tokens processed during an extended workflow. The cache read multiplier therefore has a much larger impact on multi-turn costs than the headline input price.

Price of reading 1M cached input tokens, standard tier, prompts below the long-context threshold.

What does a coding agent session cost?

To see how these rates interact, consider an illustrative coding agent session processing 8M total input tokens across approximately 50 requests of 160K tokens each. All requests stay below OpenAI's 272K-token threshold for long-context billing. The session assumes a 90% cache read rate (7.2M tokens read from cache), 0.8M tokens written to cache, and 200K total output tokens generated.

Under this workload, GPT-6 Astra generates a total cost of $27.20. That bill consists of $10.00 in cache writes (using OpenAI's 30-minute write rate of $12.50 per 1M tokens), $7.20 in cache reads, and $10.00 in output tokens. Claude Fable 5.1 runs at $21.80 for the identical session, because every one of its rates is 25% higher than on Fable 5.2.

Claude Fable 5.2 completes this session for $17.44. The cost breaks down into $8.00 in 5-minute cache writes, $1.44 in cache reads, and $8.00 in output tokens. Claude Opus 5.5 costs $9.44 for the same session, while Claude Sonnet 5.5 and GPT-6.1 Sol cost $4.72 each. Gemini 4 Argon, which has implicit caching only and no write fee, totals $4.32 at its introductory rate and $8.64 from December 1.

Comparing totals reveals that Fable 5.2 is 20.0% cheaper than Fable 5.1 and 35.9% cheaper than Astra. The Cache Gap explains why the savings against Astra exceed the 20% base price delta. Despite this drop, Fable 5.2 still bills at 1.85 times the cost of Opus 5.5 for agent workloads.

Calculated at list prices for a stated workload: 0.8M tokens written to cache, 7.2M read from cache and 200K output tokens over about 50 requests of 160K tokens. Not a measured bill.

What about one huge prompt?

Workloads that send a large standalone context without caching produce a completely different cost spread. Consider a workload consisting of a single 400K-token prompt yielding a 2K-token answer with zero cache reuse. This mirrors large document analysis, code repository ingestion or bulk data extraction.

OpenAI enforces a long-context surcharge on requests where input exceeds 272K tokens, billing the entire request at $20 per 1M input tokens and $75 per 1M output tokens for Astra. Anthropic applies no surcharge, billing Fable 5.2 at the base $8 input and $40 output rates. As a result, GPT-6 Astra costs $8.15 for this single request, while Claude Fable 5.2 costs $3.28.

Fable 5.1 costs $4.10 for the same prompt, while Claude Opus 5.5 costs $1.64. GPT-6.1 Sol reaches $1.63, Claude Sonnet 5.5 totals $0.82, and Gemini 4 Argon bills at $0.82 under introductory rates. Because OpenAI doubles Astra input rates past 272K tokens, Astra ends up costing 2.48 times as much as Fable 5.2 for this 400K-token prompt.

For standard conversational turns, the differences scale down linearly. A short chat turn of 3K input tokens and 1K output tokens costs $0.064 on Fable 5.2 compared to $0.080 on both Astra and Fable 5.1. Claude Opus 5.5 processes that same short turn for $0.032.

One 400K-token prompt with a 2K-token answer and no cache. OpenAI bills the whole request at long-context rates above 272K input tokens; Anthropic and Google charge no surcharge for these models.

Is Fable 5.2 worth twice the price of Opus 5.5?

Deciding between Fable 5.2 and Opus 5.5 requires comparing benchmark accuracy against the 1.85 cost ratio seen in agent tasks. All benchmark scores cited here are vendor-reported by Anthropic in their model release documentation. The evaluations compare Fable 5.2, Fable 5.1, Opus 5.5, and GPT-6 Astra on identical software engineering and agentic benchmarks.

On FrontierSWE v2, Fable 5.2 scores 69.9%, compared to 56.3% for Fable 5.1, 62.3% for Opus 5.5, and 65.5% for GPT-6 Astra. On OSWorld 2.1, Fable 5.2 achieves 89.3%, while Fable 5.1 records 83.5%, Opus 5.5 reaches 87.2%, and Astra scores 79.5%. On FrontierCode 2.0, Fable 5.2 reaches 49.1%, leading Opus 5.5 at 45.5%, Astra at 42.8%, and Fable 5.1 at 32.4%.

On Humanity's Last Exam with tools, Fable 5.2 scores 74.2%, ahead of Opus 5.5 at 67.7%, Fable 5.1 at 65.6%, and Astra at 57.2%. Overall, Fable 5.2 leads Opus 5.5 by 7.6 percentage points on FrontierSWE v2 and by 2.1 points on OSWorld 2.1. Engineering leads must weigh whether that 7.6-point gain on complex software tasks justifies paying 1.85 times the agent session price.

A practical routing split can balance performance and expenditure across high-volume pipelines. Developers can deploy Opus 5.5 or Sonnet 5.5 for routine turns and basic triage, reserving Fable 5.2 for the hardest coding or orchestration steps. Offline jobs can use Fable 5.2 on the Batch API at $4 input and $20 output, while fast mode should remain restricted to latency-critical user-facing paths because it doubles standard prices to $16 input and $80 output.

Methodology and limitations

All prices in this analysis reflect official list prices retrieved on November 10, 2026, from the Anthropic pricing page (platform.claude.com/docs/en/about-claude/pricing), the OpenAI pricing page (developers.openai.com/api/docs/pricing), and the Google AI Studio pricing page (aistudio.google.com/docs/pricing). Model details and benchmark scores come from Anthropic's announcement page (anthropic.com/claude-fable-and-mythos-5-2).

The workload figures represent calculated costs based on stated token distributions rather than measured production invoices. Real production sessions will vary based on prompt cache hit rates, average output length, and variance in prompt formatting. Token counts can also vary between models completing the exact same task, particularly because thinking tokens are billed at standard output rates.

Gemini 4 Carbon appears in Anthropic's published benchmark tables but does not have a public price listed on Google's AI Studio pricing page, so it was excluded from this cost comparison. All calculations exclude regional routing or data-residency uplifts, which add a 10% premium at both Anthropic and OpenAI.

Data and citation

The underlying dataset tracks every standard rate, cache read price, cache write schedule, and long-context billing rule across all compared frontier models. It includes the calculated cost of the agent session, the single 400K-token prompt and the short chat turn for each model, with a source link to each provider pricing page.

Cite this analysis as: Tokely, "Claude Fable 5.2 Pricing: What the 20% Cut Means per Task", November 10, 2026, https://tokely.me/blog/claude-fable-5-2-pricing-2026.

Originally published at tokely.me.