Google released Gemini 4 Argon (gemini-4-argon) on the Gemini API on October 13, describing it as its “most capable model built for complex coding, agentic, math, and science tasks”. Gemini API changelog The launch price of $2 per million input tokens and $10 per million output tokens runs only through November 30 and doubles on December 1, and Google’s migration notes remove features that existing Gemini code relies on. What’s new in Gemini 4 Argon
GitHub made the model generally available in Copilot the same day, which brings a separate decision for Business and Enterprise admins. GitHub changelog
A price with a seven-week shelf life
Every paid tier doubles on December 1, cached input included. Gemini API pricing
| Per 1M tokens | Through Nov 30, 2026 | From Dec 1, 2026 |
|---|---|---|
| Standard input | $2.00 | $4.00 |
| Standard output (includes thinking) | $10.00 | $20.00 |
| Standard cached input | $0.10 | $0.20 |
| Batch and Flex input / output | $1.00 / $5.00 | $2.00 / $10.00 |
| Batch and Flex cached input | $0.05 | $0.10 |
| Priority input / output | $3.60 / $18.00 | $7.20 / $36.00 |
| Priority cached input | $0.18 | $0.36 |
Every line doubles by the same factor, so a cost model built this week stays valid once you multiply the total by two. For example, an agent workload that produces 100 million output tokens a month costs $1,000 in output at the standard rate today and $2,000 from December. Any budget or customer quote built on the launch rate should show the December figure beside it.
Argon is not available on the Free tier or the Google AI Pro plan, so there is no free evaluation path through the API. What’s new in Gemini 4 Argon On the Standard and Priority tiers, grounding with Google Search and with Google Maps each includes 5,000 free requests a month, shared across Gemini 3 and newer models; after that, Search costs $14 per 1,000 requests and Maps $14 per 1,000 search queries. Gemini API pricing Google did the same with Gemini 3.8 Flash in September, with introductory API pricing that also ends in December.
Thinking tokens are billed as output, and Argon has three thinking levels: low, medium (the default) and high. Google’s own guidance is to evaluate at all three and pick the lowest level that meets the quality bar. What’s new in Gemini 4 Argon With output priced at five times input, the thinking level is the first cost setting to test.
What breaks or changes in existing code
Not every change in Google’s migration notes produces an error, so some will slip past a smoke test. What’s new in Gemini 4 Argon
| Old pattern | On Gemini 4 Argon | What to do |
|---|---|---|
Explicit caching (cachedContents, caches.create, cached_content) | Not supported | Remove it and rely on implicit caching |
temperature, top_p, top_k | Ignored if sent | Drop them; steer behavior in the prompt |
| Short HTTP client timeouts | Hard prompts can take several minutes | Set timeouts to at least 15 minutes, or use background execution |
| Client-side conversation state | Server-side state is available | Use previous_interaction_id for multi-turn and function calling |
Low max_output_tokens | The limit covers thinking and output together | Size it for both |
Implicit caching is on by default for prompts of at least 4,096 tokens and takes 95% off input tokens that hit the cache. What’s new in Gemini 4 Argon That matches the $0.10 and $0.20 cache prices above. Teams that used explicit caches for a large shared prefix lose control here: the discount now depends on requests hitting the implicit cache, so measure the hit rate instead of assuming the old savings carry over.
The ignored sampling parameters are harder to spot. A pipeline that sets temperature to 0 for repeatable extraction will keep running with no error, but the setting no longer applies. Any evaluation that depended on low-temperature output should be re-run.
Output limits and background execution
Input, thinking and output share a single window of about 1 million tokens. max_output_tokens is a hard cap on thinking plus output, and it defaults to 65,536. Without background execution, responses stop at the 64K output limit. What’s new in Gemini 4 Argon
Setting background: true lets a response run up to 1 million tokens minus the input, produced in hops of up to 64K. If you set background: true but leave max_output_tokens unset, the server still stops after 64K tokens. What’s new in Gemini 4 Argon Long generations therefore need both settings. At the December rate of $20 per million, every 100,000 output tokens in such a job costs $2.
For agent builders, the docs also list Managed Agents support through the Antigravity agent antigravity-preview-09-2026. The agent ID still carries “preview”, so treat that path as less settled than the model itself. What’s new in Gemini 4 Argon
Copilot: on by default unless an admin says otherwise
In GitHub Copilot, Argon is available on the Pro+, Max, Business and Enterprise plans. You select it in the model picker in VS Code, Visual Studio, Copilot CLI, the Copilot coding agent, the GitHub Copilot app, JetBrains IDEs, Xcode and Eclipse. GitHub says the rollout will be gradual. GitHub changelog
Usage is “billed at provider list pricing under usage-based billing”. The post does not say whether that means Google’s launch rate or the December rate, and it states no premium-request multiplier. On Business and Enterprise, admins control the model through the Copilot model policy, and under default model enablement new models are switched on automatically unless the admin has turned off the global default or disabled this model. GitHub changelog An organization that leaves the default alone has opted in, at a price that may double within weeks. Admins already working through Copilot’s October model retirements can settle the Argon policy in the same pass.
Google’s benchmark claims
Google’s model card reports 77.9% on DeepSWE v1.1, 68.9% on the Vals Index and 51.3% on AutomationBench. The same card shows Argon behind GPT-6 Astra on FrontierSWE v2, Terminal-Bench Science and OSWorld-2.0 (an offline subset, partial score), and behind Claude Opus 5.5 on Terminal-bench 4.0. Gemini 4 Argon model card These are Google’s numbers on benchmarks Google chose, and we have not found independent reproductions. The card lists a knowledge cutoff of March 2026.
What to test before December
- Run your own evaluation set at
low,mediumandhighthinking, and log output tokens per accepted result. Price the result at the December rate. - Search the codebase for
cached_content,caches.createand sampling parameters, and check how calls behave once they are removed or ignored. - Raise client timeouts, or move long jobs to background execution with an explicit
max_output_tokens. - Measure implicit-cache hit rates on prompts that share a long prefix.
- In Copilot, decide the model policy now, before usage-based charges show up on an invoice.
Still unknown: which rate Copilot passes through after December 1, how Argon’s task costs compare once thinking is counted, and whether Google’s benchmark results hold up outside its own tables.