Model A

Model B

Claude Fable 5.2Ready

The answer will appear here

Choose models and send one question below.

Gemini 4 CarbonReady

The answer will appear here

Choose models and send one question below.

Data engineering · Runnable debt reconciliation

Repair a debt reconciliation report. Deliver a complete script runnable in an empty SQLite database, not just an explanation.
Inputs: debt(country,region,principal) contains (A,R1,1000),(B,R1,2000),(C,R2,1500),(D,R2,500); country is unique. repayments(payment_id,country,amount,status) contains (p1,A,100,settled),(p2,A,50,settled),(p3,A,25,pending),(p4,B,300,settled),(p4,B,300,settled),(p5,C,200,settled); p4 is a duplicate import of one payment. benchmarks(region,year,score) contains (R1,2025,40),(R1,2026,45),(R2,2025,55),(R2,2026,60); (region,year) is unique.
The existing report joins all three tables before SUM and inflates amounts. Business rules: deduplicate payment_id; count settled only; retain countries with no repayments; join only 2026 benchmarks; outstanding=principal-paid.
Deliver: (1) One SQL block with schema, inserts and correct queries returning country detail, regional totals and global totals, sorted by country/region. Detail fields: country,region,principal,paid,outstanding,benchmark_score. (2) At least three executable validation queries detecting double counting, missing zero-payment countries and wrong benchmark years, returning PASS or FAIL. (3) Expected sample outputs and an explanation of join inflation in at most 80 words.
Keep the script compact and runnable in an empty database. Do not hard-code result rows or provide pseudocode. You have no execution tools; do not claim the script has been run.

One question, sent to both models. Each comparison starts a fresh response. Up to 3,000 input bytes. Ctrl / ⌘ + Enter to send.

Fable, Opus and GPT-6.1 Sol use low effort; Gemini 4 Carbon retains upstream defaults; Kimi uses direct-answer mode. Settings differ across vendors, so this short trial is not a controlled deep-reasoning benchmark.

Each response is billed separately and capped at 3,072 output tokens. Stopping may still incur charges.

Model comparison

Model capability positioning

These task suggestions are editorial starting points. Specifications and prices describe QVeris’s November 11 public catalog; benchmark rows retain their original settings.

Model	Official capability and evidence	Suggested first task	Usage boundary
Claude Fable 5.2	Difficult reasoning and long-horizon work. Documentation	Review a complex SQL join or break a migration into verifiable steps.	Always-on thinking. This page supplies no execution tools.
Gemini 4 Carbon	Google’s latest publicly available frontier model (November 10): extended coding, research and knowledge work. Release	Audit a multi-step debugging plan or reconcile structured financial tables.	Official agent demos include tools and repeated verification; this page is text-only.
GPT-6.1 Sol	OpenAI positions Sol for complex coding, computer use and professional work. Documentation	Review the same SQL aggregation or implement a small, testable feature.	Native tool calling requires Responses API; a text reply is not a completed agent task.
Kimi K3.1 Pro	Moonshot’s release covers coding, 3D creation and financial research. Technical blog	Try Chinese source synthesis or a parametric modeling plan.	This is Pro, not Highspeed. Tool-assisted demos require their own runtime.

Model specifications on QVeris

Model	Context tokens	Max output tokens	Catalog inputs
Claude Fable 5.2	1,000,000	128,000	file, image, text
Gemini 4 Carbon	1,048,576	943,718	image, text
GPT-6.1 Sol	1,050,000	128,000	image, text
Kimi K3.1 Pro	1,048,576	1,048,576	image, text

Catalog limits are serving limits, not a guarantee that maximum input and output fit together. File input is not enabled in this article’s composer.

Pricing and transparent usage math

Model	Input tier starts	Input Credits / M	Output Credits / M	Cache-read Credits / M
Claude Fable 5.2	0+	4,200	21,000	105
Gemini 4 Carbon	0+	1,050	5,250	52.5
GPT-6.1 Sol	0+	1,050	5,250	52.5
GPT-6.1 Sol	272000+	2,100	7,875	105
Kimi K3.1 Pro	0+	1,015.9275	5,314.1025	101.5875

Public QVeris prices checked November 11, 2026. The table preserves every published token-price tier; GPT-6.1 Sol has a second tier starting at 272,000 input tokens. No separate request fee is listed. The public catalog does not expose cache-write rates; check request-specific billing before relying on them. Live catalog

For 1,000 uncached input and 200 output tokens: cost = (input tokens × input rate + output tokens × output rate) / 1,000,000. Fable 5.2: 8.40 Credits; Gemini 4 Carbon: 2.10; GPT-6.1 Sol: 2.10; Kimi K3.1 Pro: about 2.079. This fixed-use example is not a measured bill; reasoning and longer output change usage.

Anthropic standard USD / million tokens: $8 input, $40 output, $0.20 cache read; cache writes $10 (5m) / $16 (1h). Fast mode is separately priced at $16 / $80. These vendor prices are separate from Credits; this page does not imply QVeris exposes fast mode. Official prices

Model strengths and official evaluations

Released on November 9, 2026, Claude Fable 5.2 targets difficult reasoning and extended agent work. Anthropic positions Opus 5.5 as the everyday starting point; Fable is an option when the task demands more. Official release

QVeris lists the exact model claude-fable-5-2. Its public entry supports text, images and files as input, with text output. The inline trial is text-only; it does not execute code or operate a browser.

The native model has a 1M context and 128K output limit. Adaptive thinking is always on; the native API defaults to high effort. A short page trial is not equivalent to an hours-long agent run. Model documentation

From the original release: what improved

Evaluation	Fable 5.1	Fable 5.2	Difference / setting
FrontierCode 2.0	32.4%	49.1%	+16.7 pp · Xhigh effort
FrontierSWE v2	56.3%	69.9%	+13.6 pp · Highest available effort unless noted
Chartography · no tools	44.8%	73.4%	+28.6 pp · No tools; chart recognition

Higher is better in these vendor-reported tests. They do not establish a universal win rate or predict this page’s short answers.

Official robotics evaluation: RoboDojo Sim

RoboDojo Sim tests whether a model can control a simulated two-arm robot from camera images. It probes visual understanding and action control together. Anthropic ran the models in the same harness at their highest available effort levels.

RoboDojo Sim average success rate: Fable 5.2 64.9%, Opus 5.5 39.4%, Gemini 4 Argon 34.2%, GPT-6 Astra 29.8%, Fable 5.1 27.6%.

RoboDojo average success rate (%) · highest score per model. Source: Anthropic official release. Open the chart to inspect the full-size original.

Fable 5.2 scores 64.9%, compared with Opus 5.5 at 39.4%, Gemini 4 Argon at 34.2%, GPT-6 Astra at 29.8%, and Fable 5.1 at 27.6%. The gaps are 25.5 percentage points versus Opus 5.5 and 37.3 points versus Fable 5.1.

The chart uses High effort for Gemini 4 Argon and Max for the other four models. These are vendor-reported simulation results, not QVeris reruns or a general real-world robot success rate. Argon and Astra are different models from the Carbon and Sol options above.

Independent measurements: capability and cost

Source / setting	Observed result	Interpretation
Artificial Analysis · v5.0.1 · Max / Default Fallback	Index 65; $26.54 per index task; 44.4 minutes per task.	Nine-evaluation aggregate, not 65% accuracy. These costs are not QVeris charges. Sample count is not stated in the summary.
Ultralytics · UL37 v2	360 images / 36 datasets; Fable 5.2 mAP50 0.544 vs Fable 5.1 0.427.	Default effort, production annotation pipeline plus SAM 3.1 refinement. Different from text-only reasoning tests.

Real tests: official cases and X users

The following are source-reported tests, checked November 11. Official customer examples, independent measurements and X demonstrations are labelled separately. Dollar estimates are not QVeris bills.

Official customer test: SQL correctness

Oracle’s Liyu Gong tested a query joining country debt summaries, repayment history and regional benchmarks. The comparison asks whether the business totals are correct, beyond whether the SQL runs.

Comparison	Claude Fable 5.2	Two unnamed models
Query execution	Returned a result	Ran successfully
Business totals	Matched the expected result	Overcounted totals by 5–8×
Reported outcome	Correct business result in this task	Runnable SQL produced incorrect totals

The original statement does not publish the full prompt, comparator identities or sample count; it is not a general success-rate ranking. Read the original customer statement

Community tests: inspect the actual trade-offs

Task / input	Fable 5.2 result	Comparator result	Difference / judgment	View
Voxel film-opening recreation	Fable 5.2 Extra: 29 min; $56.89 API-equivalent.	GPT-6.1 Sol xhigh: 3h26m; $12.02.	Same reference video; faster, about 4.7× the cost. Title edges and exposed set boundaries remain. No objective score.	Results / demo @thehypedotnews Test report
Three.js monster bedroom	First pass: 1h32m; $27.50 API-equivalent.	Fable 5.1: 1h14m; $28.03.	Shared room brief, different monster requirements. 5.2 is slower here; an arm pose remains awkward. Not identical inputs.	Results / demo @thehypedotnews
3D seaplane · Claude Code	Low: Lumina 3D index 1,100; High: 1,115. The X clip is Low.	Opus 5.5 xHigh: 1,097 in the author’s index.	Author prefers the scene’s aircraft, water and controls. Settings differ; index points are not success rates.	Results / demo @LuminaBench Test report
Security X-ray annotation · 10 images	Fable 5.2 refused all 10.	Fable 5.1 refused one; Opus 5.5 answered all ten.	Ultralytics reran native API requests. This is an availability failure in that dataset, not a universal vision score.	Results / demo Glenn Jocher Test report

Use these cases to choose acceptance checks: business totals for SQL, completion time and cost for agent work, visual defects for 3D output, and refusal handling for image pipelines. Complete original prompts are not available for every case.

Usage notes and FAQ

Choose two models and send the same text. Five complete tasks ask for runnable SQL with checks, an async search fix with regression tests, an interactive 3D seaplane, an auditable financial report, and a playable bedroom animation. Each includes inputs and acceptance criteria. Generated code and reports require execution or review; this page does not award automatic correctness scores for them. For the creative cases below, compare planning and code explanations here, then validate full projects in a separate execution environment.

Use the Agent comparison prompt

Click “Let an Agent compare models” at the top, then paste into your agent. The prompt includes a runnable SQL fixture and an optional 3D task, with source links, acceptance checks and cost reporting. Copying it requires no login and sends no model request. The fixtures are editorial adaptations, not the original customers’ unpublished prompts.

Is Fable always better than Opus?

No. Match the task, effort and budget. Token price alone does not predict the cost of a finished task.

How should I judge the page trial?

Check correctness before speed. Each side has a 3,072-output-token limit and no execution tools. Its results do not reproduce the original X projects.

Why can one side fail?

Availability, budget, output limits and upstream errors can differ. Preserve partial answers; an unverified model identity or incomplete response is not a zero score.

What changes in native API integration?

Adaptive thinking cannot be disabled. Omit temperature, top_p and top_k; forced tool choices any/tool are unsupported. Native documentation also specifies 30-day retention unless expressly authorized otherwise.

Claude Fable 5.2: Model Capabilities, Real Cases & Pricing Comparison