Gemini 4 Carbon
An internal Google checkpoint within the Gemini 4 model series, codenamed Carbon. It is described as a newer version than the Argon checkpoint, with reported coding capabilities comparable to those of Claude Opus 5.5.
Core Facts
Google DeepMind has released the next-generation frontier large model Gemini 4 Argon, positioned as a flagship model capable of sustained deep reasoning in complex, long-horizon professional workflows.
Google DeepMind officially released its next-generation frontier model Gemini 4 Carbon, which it calls its "strongest model."
Timeline (last 90 days)
After Gemini 4 Carbon's rate of "fabricating instead of admitting it doesn't know" rose to 11%, that rate is still the second lowest among all models.
Google did not give a specific date for opening Gemini 4 Carbon to developers and ordinary users, only saying "as soon as possible".
Third-party testing shows that the proportion of Gemini 4 Carbon forcibly fabricating answers to questions it cannot answer rose from 4% to 11%.
After personally testing the two checkpoints Carbon B and Carbon Z, the uploader concluded that their coding performance falls far short of Claude Opus 5.5 and even shows almost no meaningful improvement over existing Gemini models.
Judging by past cadence, the interval from safety personnel to Pro user access is roughly two weeks or more.
Gemini 4 Carbon ranks first globally in SWE code-fixing capability, and in SWE2 software engineering capability it is second only to the flagship models of OpenAI and Anthropic.
Gemini 4 Carbon ranks first globally in TB terminal operations capability.
Gemini 4 Carbon ranks first globally in chart analysis (BB drawing comprehension).
Gemini 4 Carbon tops both AB automated operation and VB2 financial capability.
The RB high-speed reasoning test set for Gemini 4 Carbon includes higher-order mathematical reasoning related to the Riemann hypothesis, and its performance exceeds the flagship models of OpenAI and Anthropic, ranking first globally.
40 more timeline events
All Facts (20)
Google DeepMind has released the next-generation frontier large model Gemini 4 Argon, positioned as a flagship model capable of sustained deep reasoning in complex, long-horizon professional workflows.
80%Partially VerifiedGoogle DeepMind officially released its next-generation frontier model Gemini 4 Carbon, which it calls its "strongest model."
85%UnverifiedGemini 4 Carbon is priced at $2 per million input tokens and $10 for output, with a 95% discount on cached input.
80%UnverifiedGoogle's internal tests show that Gemini 4 Carbon's success rate in generating secure code is more than twice that of the previous generation Argon.
75%UnverifiedGoogle calls Gemini 4 Carbon a "new frontier model" and emphasizes its independent evaluation performance.
70%UnverifiedThe release of Gemini 4 Carbon comes only two weeks after the release of the previous generation, Gemini 4 Argon.
70%UnverifiedGemini 4 Carbon achieved a score of 71% on the CW Bench cybersecurity test.
65%UnverifiedGemini 4 Carbon's product description is: 'Gemini 4 Carbon is our strongest model with state-of-the-art cybersecurity and long horizon capabilities.'
65%UnverifiedCarbon scored 54.8% on Exploit Gym, ranking first.
60%UnverifiedOn the Web Dev Arena leaderboard, Gemini 4 Carbon rose 89 points, from first to fourth.
60%UnverifiedGemini 4 Carbon's cheating success rate on Terminal Bench was 0.0%.
60%UnverifiedGemini 4 Carbon is first being made available to some cybersecurity defenders, and then its scope will be expanded through Fairwind.
60%UnverifiedCarbon is currently available only to trusted cybersecurity personnel, and the next batch will be distributed to some enterprise partners through a partnership program; the opening time for ordinary developers and individual users is undecided.
60%UnverifiedIn the indirect prompt injection evaluation, the attack success rate against Carbon was 0 (with 15 attempts per attack).
60%UnverifiedOn the fifth edition of the Comprehensive Intelligence leaderboard, Gemini 4 Carbon scored 59 points, ranking first and overtaking Claude Opus 5.5, 9 points higher than Argon.
60%UnverifiedIn the Grace 1 indirect prompt injection test, Gemini 4 Carbon had a 0% attack success rate across 15 attack attempts, while GPT was 8.5% and Grok was 51.8%.
60%UnverifiedGemini 4 Carbon has a vulnerability discovery rate of 88.2% and a vulnerability patching rate of 71.0%.
60%UnverifiedGemini 4 Carbon took first place on the text leaderboard in its first Arena test with a score of 1528 points.
60%UnverifiedGemini 4 Carbon's median cost per completed task is $80.
60%UnverifiedGemini 4 Carbon is currently offering preview access to enterprises and developers, with high-concurrency quotas to be gradually released in batches.
60%