Claude Fable 5.2 vs Claude Fable 5.1: Complete Comparison and Report on Pricing, Benchmarks, Effort Levels, and API Breaking Changes
Claude Fable 5.2 is the model Anthropic released on November 9, 2026 as the new top tier of its generally available lineup, ten weeks after Claude Fable 5.1 arrived on September 1, 2026. Both are the public versions of a model that Anthropic also ships as Claude Mythos for verified cyber and life-science organizations, both run with adaptive thinking that is always on, and both offer a 1M-token context window with up to 128k output tokens per request. Fable 5.1 has not been withdrawn: Anthropic's documentation lists it as "Active (legacy)" with retirement "not sooner than September 1, 2027", and it is still selectable in the Claude apps.
The verdict is clear for most buyers. Fable 5.2 costs 20% less per token ($8 input and $40 output per 1M tokens, against $10 and $50), and on every benchmark where Anthropic reports both models in the same table it scores higher, with gaps such as 49.1% vs 32.4% on FrontierCode 2.0, 74.6% vs 52.6% on Terminal-Bench-Science 0.1 and 41.3% vs 17.9% on the Remote Labor Index. Independent measurements by Artificial Analysis put Fable 5.2 at 62 on its Intelligence Index at high effort, against 47 for Fable 5.1, and show Fable 5.2 at medium effort (56) beating Fable 5.1 at its best setting (50 at xhigh) for a fraction of the cost per task.
The lower token price does not guarantee a lower bill, because Fable 5.2 spends more tokens at the same effort level: Artificial Analysis measured $9.08 per Intelligence Index task at high effort against $7.54 for Fable 5.1, and $26.54 against $12.99 at max. Teams that move to Fable 5.2 and drop one effort level get both a higher score and a lower cost; teams that keep the same effort pay more per task for a large quality gain. Fable 5.1 remains the right choice only for applications that depend on behaviors Fable 5.2 rejects with a 400 error, such as editing earlier conversation turns on older accounts, the legacy computer-use tool outside Amazon Bedrock, or a Fable 5.1 advisor. All prices in this report were verified on Anthropic's official pricing pages on November 11, 2026.
··········
CLAUDE FABLE 5.2 AND FABLE 5.1: RELEASE TIMELINE AND MODEL IDENTITY.
Anthropic describes Fable 5.2 as built "for demanding reasoning and long-horizon agentic work", the same wording its documentation applies to the Fable line in general. The company states that "Claude Fable and Claude Mythos are the same underlying model": the Fable version carries cybersecurity and biology safeguards that allow it to be offered to every customer, while Mythos 5.2 goes only to organizations accepted into the Cyber Verification Program or the Life Sciences Verification Program, plus Claude Enterprise customers through Claude Security. The same split applied to Fable 5.1 and Mythos 5.1 in September.
Fable 5.1 replaced Fable 5 at an unchanged list price of $10 and $50 per 1M tokens, with a 75% cut in cache reads to $0.25. Fable 5.2 is the first Fable release that lowers the headline token price.
........
Item | Claude Fable 5.2 | Claude Fable 5.1 |
Release date | November 9, 2026 | September 1, 2026 |
Claude API model ID | claude-fable-5-2 | claude-fable-5-1 |
Amazon Bedrock ID | anthropic.claude-fable-5-2 | anthropic.claude-fable-5-1 |
Restricted sibling | Claude Mythos 5.2 | Claude Mythos 5.1 |
Lifecycle status | Current Fable model | Active (legacy), retirement not before September 1, 2027 |
Knowledge cutoff | Not stated on the pages fetched | June 2026 (reliable and training data) |
Input and output types | Not stated separately | Text and images in, text out |
........
Artificial Analysis labels Fable 5.1 as deprecated on its model pages, but Anthropic's pricing page and release notes carry no deprecation for it, and claude.com lists it as "Fable 5.1 (legacy)" with its own price row.
··········
CLAUDE FABLE 5.2 VS FABLE 5.1 API PRICING: INPUT, OUTPUT, CACHE AND BATCH RATES.
Every line of the Fable 5.2 price list is 20% below the matching Fable 5.1 line, because Anthropic kept the same multipliers: 5-minute cache writes at 1.25 times the base input price, 1-hour cache writes at 2 times, cache reads at 0.025 times, and a 50% discount for the Batch API. Neither model has a long-context surcharge: the pricing page states that Claude 4.6 and later models, except Claude Haiku 5.5, include the full 1M-token context window at standard pricing.
........
Price per 1M tokens (USD) | Claude Fable 5.2 | Claude Fable 5.1 | Change |
Base input | $8.00 | $10.00 | -20% |
Output | $40.00 | $50.00 | -20% |
5-minute cache write | $10.00 | $12.50 | -20% |
1-hour cache write | $16.00 | $20.00 | -20% |
Cache read | $0.20 | $0.25 | -20% |
Batch input | $4.00 | $5.00 | -20% |
Batch output | $20.00 | $25.00 | -20% |
Fast mode input | $16.00 | Not offered | New |
Fast mode output | $80.00 | Not offered | New |
Long-context surcharge above 200k tokens | None | None | Unchanged |
........
Two modifiers apply equally to both models. Requests sent with `inference_geo: "us"` on the Claude API, the Claude Platform on AWS and Claude Models on Google Cloud (Anthropic managed) cost 1.1 times the standard rate in every token category, while the default global routing uses standard pricing. Fast mode, available only on Fable 5.2, doubles the base input and output prices, stacks with the caching and data-residency multipliers, and cannot be combined with the Batch API.
··········
CONTEXT WINDOW, OUTPUT LIMITS, THINKING AND EFFORT CONTROLS ON BOTH FABLE MODELS.
The two models share their core limits. Both have a 1M-token context window, which Anthropic lists as both the default and the maximum, and both return up to 128k output tokens per request. Thinking tokens count toward `max_tokens`, and the migration guide warns that on Fable 5.2 a small `max_tokens` value can end a request with `stop_reason: "max_tokens"` before any visible text is produced.
Adaptive thinking is the only thinking mode on both. Sending `thinking: {"type": "enabled"}` with a token budget, or `{"type": "disabled"}`, returns a 400 error; developers either omit the parameter or send `{"type": "adaptive"}` and steer depth with the effort setting. Forced tool use is also unavailable on both: `tool_choice` set to `any` or to a named tool returns a 400, and Anthropic directs developers to `auto` with strict tool use or to structured outputs when they need schema-valid JSON.
........
Setting | Claude Fable 5.2 | Claude Fable 5.1 |
Context window | 1M tokens | 1M tokens |
Maximum output per request | 128k tokens | 128k tokens |
Thinking mode | Adaptive, always on | Adaptive, always on |
Effort levels | low, medium, high, xhigh, max | low, medium, high, xhigh, max |
API default effort | high | high |
Default effort in Claude apps | Medium (Claude apps and Claude Code) | Medium in Claude.ai and Cowork, High in Claude Code |
Forced tool_choice (any or named tool) | 400 error | 400 error |
Computer use tool | computer_toolset_20260801 | computer_20251124 accepted |
API rate limit | Its own limit | Shared limit with Fable 5 |
Priority Tier | Not supported | Not supported |
........
The behavior of the effort setting changed even though the levels did not. Anthropic's migration guide says prompt lines such as "answer directly without deliberating" usually work on Fable 5.1 but do not stop Fable 5.2 from thinking at medium effort and above, and that at high and above the new model thinks before almost every reply. At xhigh and max inside agentic coding products, Fable 5.2 may start subagents to review its own work, which Fable 5.1 did not do on the same tasks, and the guide notes that subagents can account for a large share of session cost.
··········
FABLE 5.2 AND FABLE 5.1 ACCESS BY CLAUDE PLAN, API AND CLOUD PLATFORM.
Both Fable models sit in the same place in Claude's subscription plans. Anthropic's help center states that Fable 5.2, Fable 5.1 and Fable 5 are available on all paid plans and that the Free plan includes no Fable model. The difference between plans lies in how Fable usage is paid for, and that rule is identical for the two versions.
........
Plan or channel | Monthly price | Fable 5.2 and Fable 5.1 access |
Free | $0 | Not included |
Pro | $20 ($17 a month billed annually) | Runs on usage credits from the first request |
Max | $100 | Included up to 50% of the weekly limit, then usage credits |
Ultra | $200 | Included up to 50% of the weekly limit, then usage credits |
Team standard seat | $25 ($20 annually) | Runs on usage credits |
Team premium seat | $125 ($100 annually) | Included up to 50% of the weekly limit |
Enterprise (usage-based) | $20 per seat a month (annual, 20-seat minimum) plus API-rate usage | Billed at API rates; admins enable each model |
Claude API and cloud platforms | Per token | Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud (Anthropic and Google managed), Microsoft Foundry |
........
The launch offers apply to "every Fable model", Fable 5.1 included. Pro and standard Team seats get Fable included through November 16, 2026 at 11:59 pm PT, capped at 50% of weekly limits, plus one Fable limit reset; Max, Ultra and premium seats receive one to three resets depending on subscription length, valid until March 8, 2027. Only accounts with paid billing in place before 10:00 am PT on November 9 qualify.
Enterprise organizations see a scheduled switch. Fable 5.2 starts disabled until an admin enables it; seven days after launch it is added to every role that has any Fable model enabled, and default settings that point to a Fable model move to Fable 5.2 unless an admin opts out first. Claude Code needs version 2.1.321 or later for Fable 5.2, against 2.1.255 for Fable 5.1. For the full plan-by-plan limits, see our guide to Claude plans and usage limits.
··········
HOW ANTHROPIC'S FABLE BENCHMARKS WERE RUN AND WHERE THE NUMBERS DIFFER.
The most reliable head-to-head data comes from the benchmark table in Anthropic's Fable 5.2 announcement, which reports Fable 5.2 and Fable 5.1 side by side under the same rules. The footnotes state that Fable 5.2, Fable 5.1 and Opus 5.5 were scored with their safeguards active unless noted, and that every score uses each model's highest available effort unless noted. FrontierCode 2.0 is marked as run at xhigh for both Fable models, and the Remote Labor Index figure for Fable 5.2 is marked at high effort. RoboDojo Sim was run internally by Anthropic.
These are vendor results. Anthropic says the external organizations that tested the models, including the UK AI Security Institute and METR, did not review the blog post or the system card and do not necessarily agree with its claims. One row of the table carries Elo-style values (1785 for Fable 5.2 and 1669 for Fable 5.1) without a benchmark name, so it is left out of the comparison below.
Some Fable 5.1 numbers also moved between announcements. In September, Anthropic reported Fable 5.1 at 65.0% on Humanity's Last Exam with tools; the November table lists it at 65.6%. The September post scored OSWorld 2.0 (77.9% partial), while the November table uses OSWorld 2.1 on an offline set (83.5% partial for Fable 5.1). The comparison below uses only the November table, where both models were measured under the same version and the same rules. The table's cybersecurity and biology rows are Mythos results compared with Mythos 5.1 (for example 89.6% against 78.3% on CVE Rediscovery across 1,829 CVEs), because Fable 5.2's safeguards block most cybersecurity tasks; they describe the shared model's capability, much of which Fable customers will see blocked or answered by a fallback model.
··········
CLAUDE FABLE 5.2 VS FABLE 5.1 ON DIRECTLY COMPARABLE BENCHMARKS.
Fable 5.2 leads Fable 5.1 on all twelve comparable measurements Anthropic published. The largest gains are in long agentic work and visual or physical tasks: RoboDojo Sim more than doubles, BenchCAD 2.1 nearly doubles, and the Remote Labor Index rises from 17.9% to 41.3%. The smallest gains are on OSWorld 2.1 and Anthropic's AA-Briefcase rubric, where Fable 5.1 was already strong.
........
Benchmark (Anthropic, November 9, 2026) | Fable 5.2 | Fable 5.1 | Gap (points) | Setting |
FrontierCode 2.0 (agentic coding) | 49.1% | 32.4% | +16.7 | xhigh, both |
FrontierSWE v2 | 69.9% | 56.3% | +13.6 | Highest effort |
AutomationBench (business workflows) | 51.4% | 31.4% | +20.0 | Highest effort |
Terminal-Bench-Science 0.1 | 74.6% | 52.6% | +22.0 | Highest effort |
WANDR (relaxed score) | 82.1% | 68.7% | +13.4 | Highest effort |
OSWorld 2.1 (offline set, partial) | 89.3% | 83.5% | +5.8 | Highest effort |
RoboDojo Sim | 64.9% | 27.6% | +37.3 | Max, run by Anthropic |
AA-Briefcase v1.1 (rubric) | 66.1% | 61.5% | +4.6 | Highest effort |
Remote Labor Index | 41.3% | 17.9% | +23.4 | Fable 5.2 at high |
Humanity's Last Exam (with tools) | 74.2% | 65.6% | +8.6 | Highest effort |
Chartography (no tools) | 73.4% | 44.8% | +28.6 | Highest effort |
BenchCAD 2.1 (score 0 to 100) | 83.5 | 44.4 | +39.1 | Max |
........
At high effort, Fable 5.2 scores 45.3% on FrontierCode 2.0, which is 12.9 points above Fable 5.1's xhigh result of 32.4%. At medium effort, Fable 5.2 reaches 66.4% on Chartography, far above Fable 5.1's 44.8% at its highest setting. The company states that Fable 5.2 at medium uses fewer tokens than Fable 5.1 on challenging tasks, without giving a figure.
··········
ARTIFICIAL ANALYSIS AND EPOCH AI RESULTS FOR THE TWO FABLE MODELS.
Artificial Analysis tests each effort setting as a separate entry, which makes its pages the most useful independent source for this pair. Its Intelligence Index v5.0.1 places Fable 5.2 at max effort first of 74 tracked models with a score of 65, and Fable 5.2 at xhigh second with 62. Fable 5.1 sits between 22nd and 26th depending on effort. The Fable 5.1 entries are marked "with fallback", meaning requests blocked by safeguards were answered by Anthropic's default fallback model.
........
Artificial Analysis entry (fetched November 11, 2026) | Intelligence Index | Rank of 74 | Cost per index task | Output tokens for the index |
Fable 5.2, max | 65 | 1 | $26.54 | 275M |
Fable 5.2, xhigh | 62 | 2 | $14.83 | 144M |
Fable 5.2, high | 62 | 3 | $9.08 | 85.3M |
Fable 5.2, medium | 56 | 9 | $3.80 | 41M |
Fable 5.1, xhigh (with fallback) | 50 | 22 | $10.85 | 115M |
Fable 5.1, max | 49 | 24 | $12.99 | 163M |
Fable 5.1, high (with fallback) | 47 | 26 | $7.54 | 69.6M |
........
Fable 5.2 at medium effort scores 56, six points above the best Fable 5.1 entry, at $3.80 per task against $10.85 for Fable 5.1 at xhigh, and Fable 5.1 gained nothing from max over xhigh (49 against 50).
Epoch AI's Capabilities Index, a composite of the benchmarks Epoch tracks, gives Fable 5.2 a score of 174, first of 239 models on its model page; Epoch noted that the previous high was GPT-6 Astra at 169. The Fable 5.2 page does not show a Fable 5.1 score, so Epoch's data cannot be used for a direct gap between the two Fable versions.
··········
TOKEN CONSUMPTION AND COST PER TASK: WHY THE 20% PRICE CUT DOES NOT ALWAYS LOWER THE BILL.
The Artificial Analysis data shows that Fable 5.2 spends more tokens than Fable 5.1 at the same effort label. At high effort it used 85.3M output tokens on the index against 69.6M (22.6% more) and 7.89 billion input tokens against 5.12 billion. At max, output rose to 275M from 163M (68.7% more). The 20% price cut is smaller than these increases, so the measured cost per task went up by 20.4% at high ($9.08 against $7.54) and by 104.3% at max ($26.54 against $12.99).
Cost per index point tells a different story. Fable 5.2 at high costs about $0.146 per point of the Intelligence Index, against $0.160 for Fable 5.1 at high, $0.217 at xhigh and $0.265 at max. Fable 5.2 at medium costs about $0.068 per point, the lowest of any Fable entry. Measured by results delivered, Fable 5.2 is cheaper; measured by the invoice for an unchanged workload at an unchanged effort setting, it can be more expensive.
The practical rule follows from those figures. A team that moves from Fable 5.1 at high to Fable 5.2 at medium gains nine index points (56 against 47) and cuts the cost per task by 49.6%, from $7.54 to $3.80. A team that keeps high effort gains fifteen points and pays 20.4% more per task. Anthropic's own migration checklist recommends starting with an effort sweep at the default high setting and re-running evaluations before changing production traffic.
··········
WHAT THREE TYPICAL WORKLOADS COST ON FABLE 5.2 AND FABLE 5.1.
The examples below isolate the effect of the new price list: each workload uses the same token counts on both models, computed with the official rates verified on November 11, 2026. They do not include the extra tokens Fable 5.2 may spend at the same effort level, which the previous section measured.
The assumptions are as follows. The agentic coding session uses 8M input tokens, of which 7.2M are cache hits, 0.4M are 5-minute cache writes and 0.4M are uncached, plus 250k output tokens. The research assistant handles 1,000 requests a day, each with 30k input tokens (24k cached) and 1.5k output tokens. The document review sends one uncached 600k-token contract set and receives 20k output tokens. The batch job processes 10,000 documents of 8k input and 1k output tokens each through the Batch API.
........
Workload | Fable 5.2 | Fable 5.1 | Saving with Fable 5.2 |
Agentic coding session (8M input, 90% cached, 250k output) | $18.64 | $23.30 | $4.66 (20.0%) |
Research assistant, 1,000 requests per day | $112.80 a day | $141.00 a day | $28.20 a day (20.0%) |
Long-document review (600k input, 20k output) | $5.60 | $7.00 | $1.40 (20.0%) |
Batch job, 10,000 documents | $520.00 | $650.00 | $130.00 (20.0%) |
Same coding session in Fable 5.2 fast mode | $37.28 | Not available | Costs $13.98 more than Fable 5.1 standard |
........
With identical token counts every workload saves exactly 20%, about $846 a month for the research assistant. The fast-mode line applies the $16 and $80 rates with the standard cache multipliers on top, and doubles the cost of the session compared with Fable 5.2 at standard speed.
··········
SPEED, LATENCY AND FAST MODE ON CLAUDE FABLE 5.2.
Anthropic states that Fable 5.2 generates output up to 30% faster than Fable 5.1. Artificial Analysis measured a smaller but consistent gain in output speed: 94 tokens per second against 83 at high effort (13.3% faster), 103 against 84 at xhigh (22.6%) and 101 against 88 at max (14.8%).
Latency before the first answer token moves the other way, because Fable 5.2 thinks longer before replying. Artificial Analysis recorded 30.40 seconds to the first answer token at high effort against 13.58 seconds for Fable 5.1, 69.26 against 54.27 seconds at xhigh, and 233.95 against 165.84 seconds at max. The time to complete an average index task rose from 13.3 to 17.3 minutes at high effort, while Fable 5.2 at medium completed tasks in 7.8 minutes.
Fast mode is the new option for latency-sensitive work. Anthropic says it raises output speed by up to 2.5 times at $16 input and $80 output per 1M tokens. It is available on the Claude API without a beta header (by sending `"speed": "fast"`), in Claude Code, on Claude Models on Google Cloud in both Anthropic-managed and Google-managed form, and on Microsoft Foundry as the separate model `claude-fable-5-2-fast` on Anthropic-hosted deployments. Fable 5.1 has no fast mode on Anthropic's price list.
··········
BREAKING API CHANGES WHEN MOVING FROM FABLE 5.1 TO FABLE 5.2.
Changing the model string is the only code edit many applications need, but Anthropic lists four breaking changes that can turn working Fable 5.1 requests into 400 errors on Fable 5.2.
........
Area | Claude Fable 5.1 | Claude Fable 5.2 | What to change |
Prefix check on edited history | Enforced by default only for accounts created on or after August 31, 2026 | Enforced for every account | Stop editing system, tools or earlier turns, or use the beta drop_block behavior |
Reading thinking blocks after a model switch | Cannot read Fable 5.2 blocks (they are dropped) | Reads blocks from Fable 5.1, Opus 5.5 and other recent Claude models | Check routers and fallbacks that move conversations between models |
Computer use | computer_20251124 accepted | computer_20251124 returns 400 outside Amazon Bedrock | Move to computer_toolset_20260801 |
Advisor tool (beta) | Usable as an advisor | Fable 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as advisor | Update the advisor model ID |
Thinking blocks across accounts | Not stated | Bound to the producing account or a linked account | Avoid sharing conversation state across unlinked accounts |
........
The prefix-check error reads "The block is bound to a different conversation". As a stopgap, the beta header `thinking-binding-controls-2026-08-01` with the `drop_block` mismatch behavior lets the request proceed and lists dropped blocks in an `input_transformations` array. Accounts created before August 31, 2026 that rewrite conversation history are the most likely to hit errors, because Fable 5.1 never enforced the check for them.
Refusals on Fable 5.2 return HTTP 200 with `stop_reason: "refusal"`, every refusal counts against rate limits, and a beta server-side fallback setting can route refused requests to Claude Opus 4.8 or Claude Opus 5.
··········
SAFEGUARDS, REFUSAL FALLBACKS AND DATA RETENTION ON FABLE 5.2 AND 5.1.
Both Fable models ship with safeguards that the Mythos versions relax. Anthropic says Fable 5.2 triggers about 55% fewer blocks than Fable 5.1, with about 60% fewer in cybersecurity and about 40% fewer in biology. Blocked cybersecurity queries receive an answer from Claude Opus 4.8 and blocked biology queries from Claude Opus 5. Fable 5.1's September documentation described the same approach: it could find software vulnerabilities but not develop exploits, and penetration testing, exploit generation and binary vulnerability scanning were redirected to Opus models.
On data handling the two are identical. Both require 30-day retention and are not available under zero data retention unless Anthropic expressly authorizes it; organizations without 30-day retention receive a 400 error. Both are designated Covered Models. Anthropic is testing Enterprise Frontier Safeguards, which it says are compatible with zero data retention, in a private beta with Bridgewater, Salesforce and Stripe.
Fable 5.2 adds anti-distillation measures that Anthropic says cut the data distillers can harvest by 90% to 99%. Anthropic's alignment figures are reported for Mythos 5.2, which shares the Fable 5.2 model: an automated behavioral audit score of 2.67 against 3.28 for Mythos 5.1 (lower is better), and use of outside credentials or files in 3.1% of adversarial sandbox scenarios, against 21% for Mythos 5.1.
··········
EVIDENCE GAPS IN THE FABLE 5.2 VS 5.1 COMPARISON.
Several points remain open as of November 11, 2026. Anthropic's system-card index lists a card for Fable 5.1 and Mythos 5.1 but none for the 5.2 models at the time of checking, so the detailed safety and capability documentation for Fable 5.2 could not be reviewed. Anthropic has not published a token-usage figure behind its statement that Fable 5.2 at medium uses fewer tokens than Fable 5.1, and its migration guide gives no mapping between the two models' effort levels.
Independent coverage is still thin. Artificial Analysis offers the only per-effort comparison of both models; Epoch AI lists a Capabilities Index score for Fable 5.2 alone. No independent results were found for the agentic benchmarks where Anthropic reports the largest gains, such as the Remote Labor Index, RoboDojo Sim and FrontierCode 2.0. Fable 5.2 has no published score on Terminal-Bench 4.0, GDPval-AA v2 or CursorBench 3.2.0, the benchmarks Anthropic used to present Fable 5.1 in September.
··········
CHOOSING BETWEEN CLAUDE FABLE 5.2 AND FABLE 5.1 BY WORKLOAD.
The figures support Fable 5.2 as the default for new and existing Fable workloads, with effort chosen deliberately rather than carried over. Fable 5.1 keeps a role only where its older API behavior is still needed, and for teams that must reproduce earlier evaluation runs exactly. For a comparison of Fable 5.2 against OpenAI's flagship, see our Claude Fable 5.2 vs GPT-6 Astra report; the launch details are in our report on the Fable 5.2 and Mythos 5.2 release, and the earlier sibling comparison is in Claude Opus 5.5 vs Claude Fable 5.1.
........
Workload or situation | Recommended model | Reason |
Long agentic coding and terminal work | Fable 5.2 | +16.7 points on FrontierCode 2.0, +22.0 on Terminal-Bench-Science 0.1 |
Cost-sensitive reasoning at scale | Fable 5.2 at medium | Index score 56 at $3.80 per task, above any Fable 5.1 setting |
Highest possible quality regardless of cost | Fable 5.2 at max | Index score 65, first of 74; $26.54 per task |
Batch document processing | Fable 5.2 | Batch rates $4 and $20, 20% below Fable 5.1 |
Latency-sensitive interactive agents | Fable 5.2 fast mode or medium effort | Up to 2.5x output speed; 7.8 minutes per index task at medium against 13.3 for Fable 5.1 at high |
Code that rewrites earlier turns on accounts created before August 31, 2026 | Fable 5.1 until the code is fixed | Fable 5.2 enforces the prefix check for every account |
Computer use with computer_20251124 outside Bedrock | Fable 5.1 until migrated | Fable 5.2 returns a 400 for that tool version |
Reproducing past evaluation results | Fable 5.1 | Available until at least September 1, 2027 |
Pro or standard Team seat users | Either, on usage credits | Same billing rule; launch week included through November 16 |
........
For most API customers, the specific move the evidence supports is switching to `claude-fable-5-2`, running the effort sweep Anthropic recommends, and settling on medium or high effort: at medium, Fable 5.2 scores above every Fable 5.1 configuration Artificial Analysis tested while costing about half of Fable 5.1 at high, and at high it delivers a 15-point index gain for roughly a fifth more per task.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
