How Google Out-Iterated the Entire Model Race
Google just shipped Gemini 4 Hydrogen — its fastest, cheapest frontier-class model — two days after making Gemini 4 Carbon generally available to developers and AI Ultra subscribers. That means three Gemini 4 variants in six weeks: Argon on September 30, Carbon GA on November 10, and Hydrogen on November 12. No other lab has matched that cadence in 2026.
📖 Read the full version with charts and embedded sources on ComputeLeap →
The benchmarks are worth discussing. But they are not the story. The story is what the money says. On Polymarket, Google now holds 80% odds for best AI model at the end of November and 72% for end of 2026 — up 31% this month, the single biggest conviction move on the board. The prediction market is not reacting to any single benchmark. It is reacting to velocity.
The Timeline That Changed the Odds
Here is what Google shipped in the Gemini 4 family since late September:
| Model | Date | Role | Pricing (per 1M tokens) |
|---|---|---|---|
| Gemini 4 Argon | Sep 30 | Deep reasoning, frontier intelligence | $2 / $10 (intro) |
| Gemini 4 Carbon | Oct 28 (announced) → Nov 10 (GA) | Most complex work, cybersecurity | $2 / $10 (intro) → $4 / $20 after Nov 30 |
| Gemini 4 Hydrogen | Nov 12 | Everyday agentic workflows, speed | $1.50 / $7.50 |
Three models. Three distinct roles. Six weeks. And the pricing undercuts every competitor at the frontier tier — VentureBeat noted that Carbon's introductory rate matches Claude Sonnet 5.5 while delivering Opus-class performance, and Hydrogen at $1.50/$7.50 is cheaper still.
View the full VentureBeat analysis →
For comparison: Anthropic's Claude Fable 5.2 costs $8/$40. OpenAI's GPT-6 Astra costs $10/$50. Google is not just competing on intelligence. It is competing on the entire cost-per-task curve.
What Carbon Actually Delivers
Carbon is not a minor update over Argon. Google's own benchmarks put it first on 15 of 17 evaluations against Argon, GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1. The two exceptions: GPT-6 Astra leads on FrontierSWE 2 (software engineering), and Claude Opus 5.5 leads on OSWorld-2.1 (computer control).
View the full Google announcement →
The cybersecurity numbers stand out. Carbon leads on CWE-bench (71.0%), ExploitGym (54.8%), and SEC-bench Pro (88.2%) — all self-reported, all tested without guardrails. Whether those numbers replicate independently matters, but Google is clearly building Carbon's identity around security-forward agentic work.
On the financial side, Carbon matched the best model on Hebbia's end-to-end Financial Agentic evaluation and ranked top of 23 configurations on Balyasny's General Financial Intelligence benchmark. On Artificial Analysis' Coding Agent Index v2.0, Carbon (high reasoning, in Antigravity) ranks second at $7.76 per task — compared to $26.54 for Fable 5.2 at maximum effort.
The pricing math compounds. On Zapier's AutomationBench, Carbon scored 53.88% at $2.12 per task versus Fable 5.2's 51.45% at $4.64. Higher accuracy at less than half the cost.
Hydrogen: The Speed Play
If Carbon is Google's thinking model, Hydrogen is its execution model. Google describes it as "our daily driver for agentic workflows" — designed not for the hardest problems but for the thousands of routine tool-calling, code-generation, and file-processing tasks that make up an agent's workday.
View the full Google announcement →
The numbers that matter for Hydrogen:
- 2.5x faster than Gemini 3.8 Flash at default (high) thinking
- Intelligence Index: 51 — edging Argon's 50 and climbing to 54 at xhigh
- Terminal-Bench 4.1: 57% versus 18.2% for 3.8 Flash — a 3x improvement
- Vals Tax Agent Bench: 52.3%, ranked #1 overall
- Vals Legal Research Bench: 55.8%, tied for #1
- Native 1M-token input and up to 1M output tokens
- Four thinking levels: low, medium, high (default), and a new xhigh tier
The safety stats are noteworthy: up to 81% less agentic overreach, 12x fewer misaligned actions in tool-use evaluations, and 96.7% on an internal agent honesty evaluation. Google is clearly responding to the "agents that go rogue" narrative with hard numbers.
The Orchestrator Pattern: Carbon + Hydrogen
The most architecturally interesting signal is not either model alone. It is the Orchestrator.
In Google Antigravity, Carbon can coordinate multi-step tasks while delegating routine work to Hydrogen as a worker agent. Google reports that this pattern achieves near-Carbon task success while using about 24% less quota and finishing about 19% faster. Partners reported 2x to 2.8x faster runs with 12% to 23% lower token usage compared to 3.8 Flash alone.
This is Google shipping multi-model workflows as a first-class product — not a prompt-engineering hack, not an open-source experiment, but a productized pattern where the frontier model and the fast model collaborate within one session. If you have been building your own routing layer between expensive and cheap models, Google just made that a platform feature.
💡 The Orchestrator in practice: Carbon handles architectural decisions and complex reasoning. Hydrogen handles execution — tool calls, code generation, file processing. The result is near-Carbon intelligence at Hydrogen's speed and cost. Google reports a 24% token reduction and 19% faster completion.
What the Money Says
The Polymarket data is the sharpest signal because it aggregates thousands of bettors with real capital at risk. Here is the current state:
Model race (Chatbot Arena leaderboard):
- End of November: Google 80%, Anthropic 18%, OpenAI 2%
- End of 2026: Google 72% (+31% this month), Anthropic 21%, OpenAI 3%
Agent race:
- End of November: Anthropic 96%, Google single digits
- End of January 2027: Anthropic 75%, OpenAI 6%, Alibaba 5%
Read those numbers side by side. The market is telling us something the benchmark tables cannot: Google wins the model race. Anthropic wins the agent race. These are two different competitions.
The 31% move on the year-end model market is the largest single-month conviction shift on Polymarket's AI board. But Anthropic's 96% lock on the November agent market — built on Claude Code, computer use, and the tool-calling ecosystem — is equally striking. OpenAI is a rounding error in both.
We have been tracking prediction market intelligence on the model race throughout 2026. The trend is consistent: the crowd assigns "best model" and "best agent" to different companies, and that split has only widened.
What the Community Is Saying
The Hacker News discussion on Gemini 4 reveals the tension that benchmark tables hide. User BoorishBears called Argon "the best writing model that's been released in a while" but worried that Carbon follows "the trend of replacing promising models with benchmark-optimized versions." seizethecheese, who runs a local coding-agent benchmark, called Argon "literally the worst model in our initial ranking" — with Opus 5.5 at the top.
That split is instructive. Google's models polarize practitioners more than any other frontier lab's. The gap between Google's self-reported benchmarks and some developers' lived experience remains wider than for Anthropic or OpenAI. Whether Carbon and Hydrogen close that gap will determine whether the Polymarket odds hold.
Another pattern from the thread: multiple users found that Argon at low and medium thinking modes produced higher-quality results than high mode. If confirmed for Carbon and Hydrogen, this suggests Google's thinking models have a sweet spot below their maximum — a cost optimization that is not obvious from the benchmark tables.
⚠️ Contrarian Corner: Velocity as Weakness. Three frontier models in six weeks sounds like dominance. It might also be confusion. Developers who just integrated Argon now face a choice: stick with it, migrate to Carbon (which costs 2x after November 30), or route to Hydrogen for speed. Google's own model picker labels them "Code & analysis," "Most complex work," and "Everyday help" — but where does a medium-complexity coding task go? The surface area of decisions just tripled. Meanwhile, Anthropic shipped one model (Fable 5.2) and owns 96% of the agent market. Sometimes the fastest way to win is not to ship three models — it is to ship one that developers can actually reason about.
The Pricing War Nobody Is Talking About
Lost in the benchmark comparison is a pricing war that is reshaping the economics of frontier AI:
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Notes |
|---|---|---|---|
| Gemini 4 Hydrogen | $1.50 | $7.50 | Flat across 1M context |
| Gemini 4 Carbon (intro) | $2.00 | $10.00 | Until Nov 30 |
| Claude Sonnet 5.5 | $2.00 | $10.00 | |
| GPT-6 Sol | $2.00 | $10.00 | |
| Gemini 4 Carbon (standard) | $4.00 | $20.00 | After Nov 30 |
| Claude Opus 5.5 | $4.00 | $20.00 | |
| Claude Fable 5.2 | $8.00 | $40.00 | |
| GPT-6 Astra | $10.00 | $50.00 |
Google is offering frontier-class intelligence at commodity-class prices. Carbon during its intro period costs the same as Sonnet 5.5 — a model positioned one tier below. Hydrogen is even cheaper. If your workload involves high-volume agentic tasks, the cost differential against Fable 5.2 or Astra is 5-7x.
The catch: Carbon's intro pricing expires November 30. After that, it matches Opus 5.5 at $4/$20 — still competitive, but no longer the obvious bargain.
View the DeepLearning.ai analysis →
What This Means for You
If you are building AI-powered applications in November 2026, here is the actionable takeaway:
Test Carbon before November 30. The introductory pricing makes it the cheapest frontier model available. If your workload fits its strengths (long-horizon reasoning, cybersecurity, multimodal knowledge work), lock in your evaluation before the price doubles.
Explore the Carbon-Hydrogen Orchestrator pattern. If you have been hand-rolling a routing layer between expensive and cheap models, Google just productized that workflow. The 24% token savings with near-Carbon quality is worth benchmarking against your own routing logic.
Watch the agent market, not the model market. The Polymarket split — Google 80% on models, Anthropic 96% on agents — is the most important signal in the data. If you are building agents that need to call tools, navigate browsers, and maintain long-running workflows, the ecosystem matters more than any individual benchmark. We covered this dynamic in our earlier analysis of prediction markets and the model race.
Do not conflate benchmark leadership with developer adoption. Google leads on paper. The HN thread shows a divided community. Test with your actual workloads, not leaderboard scores. We saw the same pattern when Argon launched — strong numbers, mixed field experience.
💡 The Bottom Line: Google's iteration speed — three Gemini 4 models in six weeks — forced the prediction market's hand. But the market is smarter than the headlines: it split the "best model" crown (Google) from the "best agent" crown (Anthropic). If you are only watching one race, you are watching the wrong one.
Originally published at ComputeLeap
