TL;DR
- Carbon costs less per token: Introductory pricing is $2/$10 per million input/output tokens versus Fable 5.1’s $10/$50.
- Carbon targets complex coding: Google positions it for long-running agents and multi-step debugging.
- Cheaper tokens don’t guarantee cheaper tasks: Reasoning settings, token usage, and retries affect overall costs.
- Test before choosing: Compare both on your actual workflows, including quality, latency, safety, and cost per completed task.
- Check Fable 5.2 too: Anthropic’s newer release offers improvements and lower pricing than Fable 5.1.
Two frontier models are now competing for the same position: the brains of your coding agent. Google's Gemini 4 Carbon was unveiled on October 28 and became widely available on November 10. Claude Fable 5.1 was launched on September 1, 2026. Both are long-running, agentic work; both carry significant cybersecurity positioning, and both are priced very differently.
This comparison is for those developing products using AI agents or deciding which one to integrate into their stack. We describe the origins of each, its performance on the benchmarks that the modeler published, its cost, and how to choose among them. A word of caution: nearly all of the numbers below come from vendors, and each had a different test configuration. Use the figures as indicators, rather than judgment.
The two models at a glance
| Features | Gemini 4 Carbon | Claude Fable 5.1 |
|---|---|---|
| Maker | Google DeepMind | Anthropic |
| Announced | October 28, 2026 | September 1, 2026 |
| Broad availability | November 10, 2026 | September 1, 2026 |
| Weights | Closed | Closed |
| API price (per million tokens) | $2 input / $10 output (introductory) | $10 input / $50 output |
| Positioning | Long-horizon coding, agents, cybersecurity | Coding, science, computer use, long-running agents |
| Sibling model | Gemini 4 Argon (cheaper, earlier) | Claude Mythos 5.1 (same model, fewer safeguards, vetted access) |
In Google's Gemini 4 series, Carbon was released after Gemini 4 Argon, which was released on October 13. Fable 5.1 is part of Anthropic's Mythos series, which sits above the Opus series.
Anthropic's latest generation now uses Fable 5.2. In the Google comparison, Carbon was compared with Fable 5.1. You may need to check whether Fable 5.2 has changed your decision.
Two very different launch stories
Claude Fable 5.1 came after the earlier version was temporarily taken down, while Gemini 4 Carbon launched two weeks after Argon, initially available only to cyber warriors.
Claude Fable 5.1: A refinement after a dynamic start
Fable 5 and Mythos 5 first debuted on June 9, 2026. Anthropic suspended access to both on June 12 to comply with U.S. Department of Commerce export regulations. The Department took it offline on June 30, and Anthropic brought it back online on July 1, after about 18 days offline.
The first update to this line was Fable 5.1 and Mythos 5.1, released on September 1. According to Anthropic, they are the same model but with different safeguard levels. The generally available version is Fable 5.1, which is accessible to anyone with an API or Claude account. Mythos 5.1 is less protected and only available for vetted professionals via Anthropic's trusted-access initiatives such as Project Glasswing.
The release was described as incremental, but one number was significantly more dramatic: Terminal-Bench-Science, which tests agents on scientific problems in a terminal, went from 24.7% in Fable 5 to 52.6% in Fable 5.1.
Gemini 4 Carbon: a fast follow with a gated start
Google has announced the release of Carbon on October 27, just two weeks after Gemini 4 Argon. Google announced Gemini 4 Argon on October 13 and announced Carbon just two weeks later. Sundar Pichai put it like this: “It's a commitment to deliver models regularly.” Google described Carbon as its best offering, boasting state-of-the-art cybersecurity and long-horizon capabilities.
The roll-out was very targeted. Then, a group of cyber defenders got their first shot at Carbon via Google's Fairwind Program, and Google stated it was participating in the pre-release testing program the U.S. government has offered. The benchmarks could be read and the model not called by most developers for approximately two weeks. Most developers could read the benchmarks but not call the model for about two weeks.
This changed on November 10th. Carbon was added to Google AI Studio and the paid Google Generative AI API, to Google's AI-powered business offering, Gemini for business, and is now available to AI Ultra subscribers in the Google AI app, Google Search, and Antigravity. The same day, GitHub also announced that its Copilot feature would be available to all users. Argon is still on offer, and Google argues it's the more suitable choice for most routine use.
Coding and agentic performance
| Benchmark / Metric | Gemini 4 Carbon | Claude Fable 5.1 |
|---|---|---|
| Terminal-Bench 4.0 | 65.2% | 52.0% |
| Vals Index | Not specified | 67.87% |
| Protein binder design success rate | Not specified | 50% |
| Artificial Analysis Intelligence Index (v5.0.1) | 60 | Not specified |
| Web development leaderboard | 1,768 Elo (4th place) | Not specified |
On Terminal-Bench 4.0, Gemini 4 Carbon outperforms Claude Fable 5.1 (52.0%) and Gemini 4 Argon (57.1%). The new version, Fable 5.1, improves on the previous version, Fable 5, with a score of 42.0%. Carbon is the top-reported one, but the scores aren't directly comparable because they were evaluated in different contexts. For Fable’s own results on this test and the others Anthropic published, see our Claude Fable 5.1 benchmarks guide.
Google recognizes Carbon's prowess with long-running engineering projects like debugging gigantic codebases. It also claims it can detect 0% cheating on Terminal-Bench 4.0 and almost 80% fewer high-severity misalignment alerts in its in-house coding environment.
Coding and scientific prowess are the highlights of Fable 5.1, according to Anthropic. The model was reported to have a 50% success rate in designing protein binders, which is significantly better than the 10–20% seen at most companies. It also scored 67.87% on the Vals Index, ahead of Claude Opus 5 (67.21%) and Fable 5 (66.04%).
For Web Development, Carbon earned 1,768 Elo points, placing it fourth on the cited leaderboard.
Carbon's Artificial Analysis Intelligence Index is 60, compared to 50 for Argon and 58 for Claude Opus 5.5. These rankings include a wider performance comparison, but are not a direct comparison to Fable 5.1. For how Opus 5.5 scores across the full set of tests, see our Claude Opus 5.5 benchmarks guide.
Long context and computer use
Carbon is the most bottom-heavy. Google also outperforms Argon on GraphWalks, which evaluate reasoning over 256K to 1M tokens, with 96.2% compared to Argon's 84.2%. Google lists GPT-6 Astra at 71.8% and Claude at 66.8% in its comparison table, so they are excluded from this comparison. Fable 5.1 has a context window of one million tokens, but a published GraphWalks score is not available, so the head-to-head remains unknown.
Computer use scores are not easily comparable because the benchmark versions vary. Fable 5.1 is reported to score 77.9% on OSWorld 2.0 under partial scoring and 41.7% under strict scoring, according to Anthropic. On a sample of OSWorld 2.1 (offline), Google reported Carbon at 81.5%, while Opus 5.5 obtained 87.2%. These numbers should not be compared with other versions or subsets. More useful is that Google's own data indicates Carbon lags Anthropic's Opus 5.5 on computer use, making that a field where Carbon is not clearly leading.
Cybersecurity and safety
The companies are investing heavily in security but take different approaches.
Google reports that unguarded Carbon scores 88.2% for vulnerability detection, 54.8% for validity checking, and 71.0% for root-cause patching on SEC-bench, ExploitGym, and CWE-bench, respectively. It says Carbon more than doubled Argon's success rate in internal secure-code-generation tests.
On Gray Swan’s indirect prompt-injection benchmark, Carbon’s attack success rate was 0.0%, versus 0.7% for Argon. Google also reports reductions of about 50% in unsupported claims, 26% in overreach, and 29% in under-service.
Anthropic uses a tiered approach. Fable 5.1 retains production safeguards, which the company says now produce 60% fewer false positives. The model can identify software vulnerabilities but cannot create exploits. Vetted programs offer teams less restricted access through Mythos 5.1. In line with its EU AI Act commitment, both Fable 5.1 and Mythos 5.1 include invisible watermarks in text and file outputs.
In practice, Google's published figures come from unguarded tests, while Fable 5.1's guarded-test results are publicly available. Anthropic notes that guarded models may score lower on security benchmarks. When comparing results, check whether they come from a model you can actually use.
Pricing: The biggest practical gap
Carbon is 5 x cheaper per token than Fable 5.1 at an introductory rate. After Google's trial, Carbon is $4 and $20, both still 50% lower per token than Fable 5.1. Google has not announced an end date for Carbon's introductory pricing, but Argon's ends on November 30. Be sure to check the current rate before budgeting.
| Benchmarks | Gemini 4 Carbon | Claude Fable 5.1 |
|---|---|---|
| Input (per million tokens) | $2 introductory, then $4 | $10 |
| Output (per million tokens) | $10 introductory, then $20 | $50 |
| Caching | Cached input is 95% cheaper than fresh input | Cache reads cut 75% versus Fable 5 |
However, this is only half the price per token. Artificial Analysis found that Carbon required nearly 200 million output tokens to complete its index run, while the median across the models it tested was 73 million. Nevertheless, the price per index task was $7.76 for Carbon, compared with $10.04 for Opus 5.5. The real-world difference depends on how many tokens each model consumes for your workload.
Anthropic also didn't change the list price of Fable 5.1 from Fable 5; instead, it reduced CPU cache reads by 75%. For agentic workflows that re-read large contexts repeatedly, that change could have a bigger impact than the sticker price. According to Anthropic, they can save up to 45% on highly agentic workloads.
Availability and ecosystem
Fable 5.1 is now available on the Claude API (model ID claude-fable-5-1), Amazon Bedrock, Google Cloud, Microsoft Foundry, and in Claude's apps. In tandem, Anthropic also released updates to Claude Code, the coding chat feature, including support for separate coding sessions across different machines.
‘Carbon' is available in Google AI Studio, Gemini for business, the Gemini app (for AI Ultra subscribers), Search's AI Mode and Antigravity, and GitHub Copilot. Unlike Fable 5.1, individual users won't hit a paywall for Carbon access, as it will be available only at the top-tier subscription.
Both models are closed-weight and API-only, meaning that if you are looking to self-host, neither model is an option.
What builders must check before taking Gemini Carbon 4
1. Don't pick on benchmarks alone: All major results reported above are self-reported and under various harnesses. Test both models on a set of your actual jobs: your repos, prompts, tool configuration. Track success rate, latency and overall cost per completed task in terms of tasks, not tokens.
2. Cost-sensitive agent loops favor Carbon on paper: If you have a long-running product with lots of agents, you should definitely give this model a shot if it costs a bit more than Fable 5.1 per token but doesn't increase token usage.
3. Quality-critical and regulated work favors a careful look at Fable 5.1: For some teams, the tiered safeguards and published improvements to refuse, plus the company's more extensive production history, may be more important than the actual price. There's also been a bit more independent testing for Fable 5.1, which has been available for more than two months.
4. Routing plan, not loyalty: Even Google labels Argon as the correct option for difficult problems, with Carbon for only hard problems. The same goes for the vendors: a cheap model for the easy steps, a frontier model for the harder ones. Design your product to allow a model change, not a rewrite.
5. Watch the release cadence: Google released two frontier models in two weeks, and Anthropic has leaped from 5.1 to 5.2. Whatever you choose to do today will only last a short time.
The verdict
There is no winning side, and anyone who says otherwise is telling you what they think the vendor wants you to hear.
Gemini 4 Carbon has the most published evidence of agentic terminal coding, long-context retrieval, and pricing. It's also new, and its record has yet to be written. Claude Fable 5.1 is older, ranks highly on the Vals Index, has more scientific/agentic tasks, and offers a safeguard model with a higher token price that lets teams access it as either a guarded or vetted model.
Try out Carbon for pricing and long-horizon coding first if that's your need. If a longer production record is desired and an Anthropic safeguard approach is preferred, it is better to go with Fable 5.1 first and determine whether Fable 5.2 is a better goal.
However, you might need a platform to build your own software. Emergent helps you build your agent or application and run your business as you choose. Now, you can develop and deploy coding agents using artificial intelligence through Emergent. Test advanced models, streamline development workflows, incorporate intelligent software, and turn your ideas into applications without manual coding.

Describe it. Build it. Launch it.
- End-to-End Creation
- No Code Needed
- Live in Minutes

Most AI app builders stop at prototypes. Emergent creates production-ready apps you can actually launch.
- Production-ready apps
- Web & mobile apps
- Deploy in minutes






