Gemini 4 Carbon vs Claude Fable 5.2 vs GPT-6.2 Astra: Which Hard-Task Model Is Worth Paying For?
Compare Gemini 4 Carbon, Claude Fable 5.2 and GPT-6.2 Astra on official prices, limits, effort controls, Google MKA results and Artificial Analysis cost per task, then choose by task.
Gemini 4 Carbon is the cheapest way to get close to Claude Fable 5.2 on long, file-heavy knowledge work. On Google DeepMind’s MKA benchmark, Fable 5.2 scored 2.22 points higher, a gap Google’s own bootstrap test could not separate from noise, at 6.26 times Carbon’s cost per task (with Carbon priced at its December rate). That cost advantage does not hold against GPT-6.2 Astra. On Artificial Analysis’s broad Intelligence Index, Astra at High effort matches Carbon at High for about half of Carbon’s current-rate cost and in a fraction of the time. Pay for Fable 5.2 when the last few points of quality matter more than a cost per MKA task that is 5–6 times Carbon’s (Carbon at December rates). Pick Astra when you need speed and good scores per dollar. Pick Carbon for visually dense documents, very long context and multi-hour agent runs. Its token rates double on December 1, 2026, so budget at the new rate.
This comparison is current as of November 19, 2026. Ofox publishes it, and two of the three models, Claude Fable 5.2 and GPT-6.2 Astra, can be called through Ofox. Gemini 4 Carbon cannot. We checked Ofox’s /v1/models list on November 19: it returned 137 models, including anthropic/claude-fable-5.2 and openai/gpt-6.2-astra, but no gemini-4-carbon. Carbon examples below therefore use Google’s own Gemini API.
Each kind of evidence is labelled where it appears. Doc facts come from the vendors’ current pages. Benchmark results come from Google DeepMind’s MKA paper and from Artificial Analysis (AA), fetched on November 19. Our calculations are arithmetic on those numbers. Our runs are four illustrative requests through Ofox, not a benchmark.
One trap before any comparison: whose model versions are in the table?
Most launch charts circulating for these models compare a different set of versions from the one you’d buy today:
| Source | Carbon compared with | Not included |
|---|---|---|
| Google’s Carbon model card (Nov 10) | GPT-6 Astra, Claude Opus 5.5, Claude Fable 5.1 | Fable 5.2, GPT-6.2 Astra |
| Google DeepMind’s MKA technical paper (Nov 10) | Claude Fable 5.2, Opus 5.5, GPT-6 Astra | GPT-6.2 Astra |
| Anthropic’s Fable 5.2 release (Nov 9) | Fable 5.1, Opus 5.5, GPT-6 Astra, Gemini 4 Carbon | GPT-6.2 Astra |
| Artificial Analysis MKA leaderboard (live) | All three current versions | — |
GPT-6.2 Astra shipped on November 17 (OpenAI API changelog), after every vendor table above was published. A score labelled “GPT-6 Astra” is not a GPT-6.2 Astra score. A “Fable 5.1” column on the Carbon card says nothing about Fable 5.2. This article keeps those labels exactly as each source printed them. The only place all three current models appear under one harness is Artificial Analysis.
For the earlier generation, see our Gemini 4 Argon vs GPT-6 Astra and Claude Opus benchmark guide and the GPT-6 Astra review. If you are weighing Fable 5.2 against Anthropic’s own models, read Fable 5.2 vs 5.1 vs Opus 5.5.
Official prices on the same basis
Doc facts, Standard tier, USD per 1 million tokens, checked November 19, 2026:
| Gemini 4 Carbon | Claude Fable 5.2 | GPT-6.2 Astra | |
|---|---|---|---|
| Model ID | gemini-4-carbon | claude-fable-5-2 | gpt-6.2-astra |
| Input | $2.00 through Nov 30; $4.00 from Dec 1 | $8.00 | $8.00 |
| Cached input (read) | $0.10; $0.20 from Dec 1 (implicit caching only) | $0.20 (2.5% of input) | $0.40 |
| Output (thinking included) | $10.00; $20.00 from Dec 1 | $40.00 | $40.00 |
| Discount options | Batch, Flex and Priority tiers (not priced here) | Batch 50% off; fast mode costs more ($16 / $80) | Background windows 30% (2 h) or 40% (8 h) off; Batch |
| Long-context surcharge | None; flat across the 1M window | None; standard rate across the 1M window | Above 272K input tokens: input and cache rates double, output ×1.5, for the whole request |
| Free tier | None | None | None listed |
Sources: Google’s Carbon pricing section and Carbon guide, Anthropic’s models overview and Fable 5.2 release, OpenAI’s GPT-6.2 Astra model page and changelog. Cache writes are billed separately for Fable and Astra, and Google also bills grounding tools per query. Our December price-change guide covers the Carbon and Argon rate change in detail.
Fable 5.2 and GPT-6.2 Astra share the same list rate below 272K input tokens, so their difference comes down to how many tokens each spends. Above that, Astra’s long-context surcharge makes it the most expensive per token of the three, which matters for the 256K–1M workloads where Carbon is strongest. Carbon’s list price is one-quarter of theirs until November 30 and one-half from December 1.
Our calculation: one identical request. Take 50,000 input tokens, 40,000 of them cache hits, and 6,000 output tokens. Cache-write fees are ignored, and Carbon only offers implicit caching, so its hit rate is not guaranteed.
| Uncached input | Cached input | Output | Total | |
|---|---|---|---|---|
| Carbon (Nov rates) | $0.020 | $0.004 | $0.060 | $0.084 |
| Carbon (Dec rates) | $0.040 | $0.008 | $0.120 | $0.168 |
| Fable 5.2 | $0.080 | $0.008 | $0.240 | $0.328 |
| GPT-6.2 Astra | $0.080 | $0.016 | $0.240 | $0.336 |
This table is the least useful one in the article. In practice the three models do not spend 6,000 output tokens on the same task. In Google’s MKA paper, Fable 5.2 averaged 378.9k output tokens per task, Carbon 129.5k and GPT-6 Astra 42.0k. Tokenizers differ too: in our runs, the same prompt counted as 301 input tokens for Fable 5.2 and 212 for GPT-6.2 Astra. Price per token sets the rate, but cost per completed task decides the bill. For a full Carbon task ledger with retries and cache rules, use our Carbon task-cost calculator.
Limits, effort controls and access
Doc facts:
| Gemini 4 Carbon | Claude Fable 5.2 | GPT-6.2 Astra | |
|---|---|---|---|
| Context window | 1,048,576 tokens, shared by input, thinking and output | 1M tokens | 1,050,000 tokens (max input 922,000) |
| Output per response | 65,536 synchronous; up to 1M minus input with background long decode | 128K | 128,000 |
| Reasoning control | thinking_level: low, medium (default), high | output_config.effort: low to max, default high; adaptive thinking always on | reasoning.effort (reasoning_effort in Chat Completions); none is rejected; the model page does not list the other levels or the default |
| Input modalities | Text, image, audio, video | Text, image (plus PDF via the API) | Text, image |
| Knowledge cutoff | March 2026 | June 2026 (reliable) | May 18, 2026 |
| Where to call it | Gemini API paid tier, Google AI Studio, Gemini for business | Claude API, Bedrock, Google Cloud, Foundry | Chat Completions, Responses (needed for tool calling), Batch |
| On Ofox (checked Nov 19) | No | Yes | Yes |
Some details matter before a migration:
- Carbon’s output cap includes its thinking.
max_output_tokenscovers thinking plus answer, and Google recommends at least 32,768 tokens because Carbon spends more of its limit on thinking than Argon did. Longer outputs needbackground: true. Our Carbon API setup guide walks through both. - Fable 5.2 cannot switch off thinking. Anthropic recommends starting at
high, usingxhighfor the hardest long-running work, and keepingmaxfor measuring the model’s ceiling. Our Fable 5.2 API guide covers request shape and history handling. - GPT-6.2 Astra’s tool calling requires the Responses API. The changelog also flags migration changes in how Responses keeps reasoning after edited history and in how it treats the top-level
instructionsfield.
On Ofox, the Fable 5.2 model page lists both OpenAI-compatible and Anthropic protocols. Current Ofox prices for both models are shown on their model pages and on the pricing and billing docs; we don’t restate them here.
Benchmark 1: Google DeepMind’s MKA paper (the “statistical tie”)
The claim that Carbon “ties Fable 5.2 at a sixth of the cost” circulates mostly through secondary coverage. Its primary source is the MKA technical paper. Google’s MKA blog post only says Carbon ranked “#2 in quality” and completed tasks “at 6x lower cost on average” than the top model.
MKA uses 200 private tasks across 10 industries. Each is a multi-step project built on visually dense files: drawings, schematics, scans and charts. Every model ran in the same open-source Stirrup harness with a 250-turn limit. The paper’s main tables:
| Model (as named in the paper) | Climb-WSR (± 95% CI) | Overall WSR | Output tokens / task | Turns | Time / task | API cost / task |
|---|---|---|---|---|---|---|
| Claude Fable 5.2 | 56.56 ± 3.63 | 73.26 | 378.9k | 128.8 | 83.2 min | $35.28 |
| Gemini 4 Carbon | 54.34 ± 3.24 | 71.89 | 129.5k | 46.6 | 29.2 min | $5.63 |
| Claude Opus 5.5 | 53.01 ± 3.54 | 70.98 | 208.8k | 91.4 | 33.8 min | $10.29 |
| GPT-6 Astra | 50.24 ± 3.25 | 69.34 | 42.0k | 29.0 | 18.5 min | $6.03 |
The paper words its central finding this way: under a paired task-level bootstrap, Fable 5.2, Carbon and Opus 5.5 “are statistically indistinguishable on Climb-WSR”. Fable 5.2’s lead over Carbon is +2.22 points with a 95% interval of −0.94 to +5.49. Carbon’s lead over GPT-6 Astra (+4.10 points, interval +1.00 to +7.11) is significant. The 6.26× cost gap comes from $35.28 ÷ $5.63 at “provider list prices at the time of writing”. The paper doesn’t say which Carbon rate it used, but our arithmetic points to the December rate: at $4 / $20, Table 4’s Carbon token counts give $5.63 when 93.7% of input is served from cache, which is the bottom of the paper’s stated 93.7%–97.7% range. At $2 / $10 the same tokens cannot reach $5.63 at any cache share in that range. At today’s November rate, Carbon’s cost per task would be about half, and Fable’s multiple about twice as large.
Read it with four limits in mind:
- Each model ran once per task, graded in a single pass. The paper says its intervals capture task sampling, not run-to-run or grader variance.
- Google co-designed the benchmark and its model came second. The grader is a non-Google model, and humans agreed with it on 96.2% of the high-weight criteria they checked. That is a strong check, but not an independent replication.
- The OpenAI model tested was GPT-6 Astra, not GPT-6.2 Astra. The paper says nothing about 6.2.
- A statistical tie is not a guarantee of equal results. It means this 200-task sample cannot rule out that the true gap is zero. It equally cannot rule out a gap of five points in Fable’s favor.
Benchmark 2: the Artificial Analysis MKA leaderboard, now with GPT-6.2 Astra
Google built MKA together with Artificial Analysis, and AA hosts the public leaderboard. For the four models in the paper, AA’s Climb WSR scores match the paper to two decimals. Treat the leaderboard as the same evaluation extended to newer models, not as an independent replication. AA’s own judge is GPT-6 Luna at medium effort, and the effort level is part of each model name. AA estimates cost and time on its own basis, so those columns differ from the paper’s. Fetched November 19, 2026:
| Model and effort (AA labels) | MKA Climb WSR | Cost per task | Time per task |
|---|---|---|---|
| Claude Fable 5.2 (Max) | 56.6% | $31.98 | 57.1 min |
| Gemini 4 Carbon (High) | 54.3% | $6.29 | 29.0 min |
| Claude Opus 5.5 (Max) | 53.0% | $10.23 | 26.1 min |
| GPT-6.2 Astra (Max) | 52.7% | $3.67 | 21.8 min |
| GPT-6 Astra (Max) | 50.2% | $6.21 | 17.0 min |
Two things change once GPT-6.2 Astra is on the board:
- On cost, Fable 5.2 to Carbon is about 5.1× ($31.98 ÷ $6.29, our calculation) on AA’s figures, rather than the paper’s 6.26×. The direction holds. The size depends on how each party prices the same token usage.
- GPT-6.2 Astra is now the cheapest of the three on MKA, 1.7 points below Carbon (52.67% vs 54.34%) at roughly 58% of Carbon’s AA cost. AA publishes no confidence intervals, and the paper never tested 6.2, so nobody has said whether that 1.7-point gap is real.
Check the price basis before reusing AA’s Carbon figures. AA’s data prices Carbon at $4 / $20, the December rate. Every Carbon token rate exactly doubles on December 1, so at today’s $2 / $10 the same token usage would cost about $3.15 per MKA task (our calculation: $6.29 ÷ 2). That puts Carbon slightly below GPT-6.2 Astra’s $3.67 until November 30, and back at $6.29 afterwards.
Gemini 4 Neon at Xhigh effort, released on November 18, scored 56.5% at $5.36 on the same board. Neon is a separate comparison, covered in Gemini 4 Neon vs Carbon.
Benchmark 3: the broad Artificial Analysis Intelligence Index
MKA measures one kind of work. AA’s Intelligence Index v5.0.1 combines nine evaluations covering knowledge work, coding, science and general knowledge. Values from AA’s model pages, fetched November 19, 2026, with AA’s cost at its listed price for each model:
| Model and effort | Index score | Cost per index task | Time per index task |
|---|---|---|---|
| Claude Fable 5.2 (Max) | 64.6 | $26.54 | 41.6 min |
| Claude Fable 5.2 (High, the API default) | 61.8 | $9.08 | 15.0 min |
| GPT-6.2 Astra (Max) | 62.1 | $4.00 | 13.7 min |
| GPT-6.2 Astra (High) | 60.1 | $2.07 | 5.9 min |
| Gemini 4 Carbon (High) | 60.0 | $7.76 (≈ $3.88 at Nov rates) | 37.2 min |
| Gemini 4 Carbon (Medium, the API default) | 58.3 | $6.59 (≈ $3.29 at Nov rates) | 29.6 min |
The halved Carbon figures are our calculation, and the time column converts AA’s seconds to minutes. On this broader mix, Carbon’s cost story reverses. GPT-6.2 Astra at High matches Carbon at High (60.1 vs 60.0) for about half of Carbon’s current-rate cost and a quarter of AA’s December-rate cost, in about one-sixth of the time (5.9 vs 37.2 minutes). AA’s Carbon page also lists a 73-second time to first answer token and an output speed of 55 tokens per second, and calls the model “notably slow per task and very verbose”. At each vendor’s default, Fable 5.2 High (61.8) outscores Carbon Medium (58.3) and costs $9.08 per task against Carbon’s $6.59 at the December rate. OpenAI doesn’t publish Astra’s default effort, so we compare it at High.
What vendor tables add (and where they don’t apply)
Anthropic’s Fable 5.2 release is the only first-party table that prints Fable 5.2 and Carbon side by side. Its OpenAI column is GPT-6 Astra. Anthropic used each model’s highest effort unless noted, and marked Carbon’s FrontierCode score as Medium effort.
| Benchmark (Anthropic-reported) | Fable 5.2 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|
| FrontierSWE v2 | 69.9% | 65.5% | 61.2% |
| AutomationBench | 51.4% | 41.4% | 53.9% |
| Terminal-Bench-Science 0.1 | 74.6% | 68.1% | 64.8% |
| AA-Briefcase v1.1 (Elo) | 1785 | 1569 | 1611 |
| Chartography (no tools) | 73.4% | 71.0% | 72.5% |
Carbon’s own card adds long-context evidence that no other table covers: on GraphWalks BFS from 256k to 1M tokens, Google reports Carbon at 96.2% against 71.8% for GPT-6 Astra and 65.0% for Fable 5.1. Fable 5.2 was not measured there, so we treat Carbon’s long-context lead over Fable 5.2 as plausible but unverified. The same card puts Carbon at 65.2% on Terminal-Bench 4.0 (AA), against 59.1% for GPT-6 Astra and 52.0% for Fable 5.1, and at 9.6% on SWE-sweep against 9.5% and 2.9%. Across these sources, Fable 5.2 leads agentic coding and open-ended research, and Carbon is strongest on business workflows and long, visual documents. GPT-6 Astra beats Carbon on FrontierSWE v2 and Terminal-Bench-Science in Anthropic’s table, but trails it on AutomationBench, Chartography and Google’s long-context test.
Our illustrative runs: Fable 5.2 vs GPT-6.2 Astra on Ofox
We wanted to see the token-spend difference on our own prompts, so we ran two small tasks on November 19, 2026. Each model got each task once, with identical prompts, through Ofox’s OpenAI-compatible endpoint (https://api.ofox.ai/v1/chat/completions), with reasoning_effort: "high" and max_tokens: 32000. Carbon was not run because Ofox does not offer it. Costs below apply each vendor’s official list price to the returned usage. They are not Ofox charges.
- Task A, code fix: a 9-line interval-merging function that violates a written spec four ways (mutates input, keeps invalid intervals, misses touching intervals, returns lists not tuples). We checked each corrected function against 505 test cases, including mutation and return-type checks.
- Task B, counting puzzle: count the orderings of seven items under five constraints, without code. A brute-force check gives 129.
| Task | Model | Result | Output tokens (reasoning) | Wall time | List-price cost |
|---|---|---|---|---|---|
| A | Claude Fable 5.2 | All 4 issues found; tests pass | 1,037 (465) | 12.1 s | $0.044 |
| A | GPT-6.2 Astra | All 4 issues found; tests pass | 201 (49) | 5.1 s | $0.010 |
| B | Claude Fable 5.2 | 129, correct | 5,250 (4,419) | 44.7 s | $0.211 |
| B | GPT-6.2 Astra | 129, correct | 2,435 (1,916) | 29.8 s | $0.098 |
Both models solved both tasks. Fable 5.2 used 5.2× and 2.2× as many output tokens and wrote fuller explanations, including a worked example for Task A. Astra answered tersely. At equal list prices, the difference in token spend became the whole cost difference. That matches the direction of the AA and MKA token counts, but four requests prove nothing about quality or stability. We also did not verify how Ofox maps reasoning_effort onto Fable’s effort setting on the OpenAI-compatible route. The raw requests, responses and verifier are kept with our report files.
To reproduce this through one key, point an OpenAI-compatible client at Ofox and change only the model ID. Setup details are in the Ofox integration docs.
from openai import OpenAI
client = OpenAI(base_url="https://api.ofox.ai/v1", api_key="YOUR_OFOX_KEY")
for model in ["anthropic/claude-fable-5.2", "openai/gpt-6.2-astra"]:
r = client.chat.completions.create(
model=model,
max_tokens=32000,
reasoning_effort="high",
messages=[{"role": "user", "content": "<your task>"}],
)
print(model, r.usage.completion_tokens, r.choices[0].message.content[:200])
Carbon has to be tested on Google’s native API with a Gemini API key. This is the documented Interactions call:
from google import genai
client = genai.Client()
interaction = client.interactions.create(
model="gemini-4-carbon",
input="<your task>",
generation_config={"thinking_level": "medium"},
)
print(interaction.output_text)
If a request fails, separate access and quota errors from parameter errors before retrying. For Fable, our Fable 5.2 400-error guide covers thinking and tool-history mistakes. For Astra-family model ID, 401 and token-limit errors, see the GPT-6 Astra error guide. For the Gemini side, see Gemini 4 API errors. Ofox’s error-handling guide and rate-limit docs cover retries and limits on the Ofox route.
Decision table: which one for which job
| Your situation | Choose | Why (evidence) | Watch out for |
|---|---|---|---|
| Long reports, slide decks or audits built from drawings, scans and charts | Gemini 4 Carbon | Statistically tied with Fable 5.2 on MKA at about 1/6 (paper) to 1/5 (AA) of its cost per task, both at Carbon’s December rate | Slow; doubles on Dec 1; not on Ofox |
| Inputs of 256k–1M tokens that must be reasoned over, not just retrieved | Gemini 4 Carbon | 96.2% on GraphWalks 256k–1M on Google’s card | Fable 5.2 not measured there; check yourself |
| Business automation across SaaS tools | Gemini 4 Carbon | 53.9% on AutomationBench vs 51.4% for Fable 5.2 (Anthropic-reported) | Small gap, one benchmark |
| Hardest agentic coding, migrations, open research where a few points matter | Claude Fable 5.2 | Leads FrontierSWE v2, Terminal-Bench-Science and AA Index at Max | MKA cost per task was 5–6× Carbon (Carbon at December rates); Max effort is slow |
| Work where you need the highest single score and cost is secondary | Claude Fable 5.2 (Xhigh/Max) | Top AA Index (64.6) and top MKA score | $26.54 per index task at Max; try High first |
| Security research | Not Fable 5.2 on default access | Anthropic says Fable 5.2’s safeguards block most cybersecurity tasks; its Mythos 5.2 needs Cyber Verification Program access | Google also offers Carbon to verified defenders through its Fairwind Program; we did not compare cyber results here |
| Interactive or latency-sensitive hard reasoning | GPT-6.2 Astra | Fastest of the three on AA (5.9 min per index task at High) and cheapest at equal scores | New; few independent evals yet; no audio or video input |
| You need one API key across vendors, with fallback | Fable 5.2 + GPT-6.2 Astra on Ofox | Both available behind one endpoint | Carbon would need a separate Google account |
What each one is worse at
- Gemini 4 Carbon: slow and wordy (37.2 min per AA index task, 73 s to first answer token). Its Intelligence Index score per dollar is no better than GPT-6.2 Astra’s. No free tier, the token price doubles on December 1, its knowledge cutoff (March 2026) is the earliest of the three, and it isn’t on Ofox.
- Claude Fable 5.2: by far the most expensive per finished task on MKA, in both the paper’s and AA’s cost figures, because it spends far more tokens (40.48M total tokens per MKA task vs 7.05M for Carbon). Thinking cannot be turned off, and cybersecurity work is mostly blocked.
- GPT-6.2 Astra: released November 17, so most vendor tables only cover GPT-6 Astra. It scored lowest of the three on AA’s MKA leaderboard, accepts only text and image input, and needs the Responses API for tool calling.
How to settle it for your own workload
Benchmarks narrow the shortlist, but your own tasks decide it. Choose 20–50 real tasks that you can grade, keep the prompts identical, and record score, total tokens, wall time and list-price cost for each attempt, including failed ones. Run Carbon at medium and high, Fable 5.2 at high before xhigh, and GPT-6.2 Astra at high before max. If you’re testing Carbon this month, compute its cost at both $2/$10 and $4/$20, so a November trial doesn’t set a budget that December breaks.
If two models land within a few points of each other and one costs five times as much, the MKA data says the gap may be noise. For a broader view of the three vendors’ tiers, see Gemini 4 vs Claude vs GPT model tiers and the Gemini 4 Neon vs Claude Opus 5.5 vs GPT-6.2 Sol comparison one tier down.
Related Docs
- Ofox integration guides for connecting OpenAI- and Anthropic-compatible clients
- Ofox Gemini API reference for the Gemini models Ofox does list (Carbon is not among them)
- Error handling and retries
- Rate limits
- Ofox pricing and billing docs


