Anthropic has been setting the pace for AI for Code since the start of the year. With Gemini 3.5 Flash and a completely redesigned Antigravity, Google is finally trying to get things back together. The speed and price are there.
With Gemini 3.5 Flash, Google is advancing its pawns in the highly competitive sector of AI for code. Less expensive, as powerful and up to five times faster than Claude Opus 4.7 or GPT-5.5, the model presented by DeepMind has everything to please developers. Add to that a redesigned experience in Antigravity and a CLI mode, and Google is openly hunting in Anthropic territory. Review of the forces present.
The real advantages of Gemini 3.5 Flash
With the arrival of Gemini 3.5 Flash, Google promises that it is no longer “necessary to sacrifice quality for latency.” The model achieves a tour de force in the new race of AI for code: that of speed. Gemini 3.5 Flash is the fastest code model Google has released to date. According to data from Artificial Analysis, from Google AI Studio, the model reaches a speed of 277 tokens per second, compared to 65 tokens per second for GPT-5.5 (xhigh mode) and 51 tokens per second for Claude Opus 4.7 (Max reasoning). It is therefore approximately 4.3 times faster than GPT-5.5 and 5.4 times faster than Claude Opus 4.7. A concrete difference on a daily basis for companies seeking to produce code ever faster, and for agentic systems in particular.
The feat is real, especially since the model is not significantly worse than the competition’s frontier models. On agentic tasks, Gemini 3.5 Flash even takes the lead with 83.6% on MCP Atlas (multi-step workflows orchestrated via the MCP protocol), ahead of the 79.1% of Opus 4.7 and the 75.3% of GPT-5.5. On Terminal-bench 2.1 (code in terminal environment), its 76.2% clearly exceeds Opus (66.1%) and passes just below GPT-5.5 (78.2%).
The interface check confirms, once again, parity: on OSWorld-Verified (computer use), Flash reached 78.4%, neck and neck with GPT-5.5 (78.7%) and Opus (78.0%). Only dropout: SWE-Bench Pro (varied agentic code tasks). With 55.1%, Gemini 3.5 Flash remains behind the 64.3% of Opus 4.7 and the 58.6% of GPT-5.5. In other words, Flash holds the comparison everywhere and often wins in agentic, except when solving complex code tasks independently. The most demanding projects will still have to go through Claude or GPT-5.5… while waiting for Gemini 3.5 Pro?
The real question mark: the scaffold
Why has Google never really taken the lead on AI for code? Until now, everything was in the scaffold. The entire software layer around the model matters as much, if not more, than the raw model itself. An excellent model may exhibit poor code performance if the scaffold is not optimal. Anthropic understands this well and today takes as much care of Claude Code (the scaffold) as its raw models. The results are there: the reference for the developer community, including in France, remains Claude Code.
To try to reverse the trend, Google is offering a brand new version of its Antigravity code agent. The entire experience has been redesigned for developers based on extensive field feedback. Antigravity 2.0 no longer has anything to do with the heavy and barely usable IDE offered until now by Google. The tool allows you to orchestrate several code agents in parallel, like Claude Code in Claude Desktop. But the real change comes with the announcement of Antigravity CLI. Google is banking almost everything on access to the agent via the command line, widely acclaimed by developers with Claude Code. Gemini CLI will also migrate to Antigravity CLI. The agent allows you to carry out the majority of tasks handled by Claude Code (goal, routine, etc.). Will this be enough to cannibalize Anthropic users? Difficult to say, but to stay the course, Google will have to follow the shattering pace imposed by Anthropic since the beginning of the year on the code, while offering a truly effective agent (test to come on the JDN).
Google pricing
With $1.50 in input and $9 in output per million tokens, Gemini 3.5 Flash is by far the cheapest of the three. In terms of input, it costs 3.3 times less than Claude Opus 4.7 and GPT-5.5, both at $5. In terms of output, the gap narrows but remains clear: 2.8 times less expensive than Opus 4.7 ($25) and 3.3 times less expensive than GPT-5.5 ($30). Google pricing, as always. The company controls the entire chain, from hardware to model, thanks to its in-house TPUs (Google is exempt from the Nvidia tax).
| Model | Input (for 1M tokens) | Output (for 1M tokens) | |
|---|---|---|---|
| Claude Opus 4.7 | $5 | $25 | |
| Gemini 3.5 Flash | $1.50 | $9 | |
| GPT-5.5 | $5 | $30 | |
For developers who would use Google AI subscriptions directly, Google offers four packages in France: AI Plus at 7.99 euros per month, AI Pro at 21.99 euros per month, AI Ultra (new) at 99.99 euros per month and AI Ultra x5 at 219 euros per month. Prices align with OpenAI and Anthropic pricing policy. Even if Google, like its competitors, does not communicate on the number of tokens included, the pricing of the API strongly suggests that the same level of package (eg: 21.99 euros) gives the right to significantly more generous use at Google.
Should you upgrade to Gemini 3.5 Flash?
As with every time a new model or major update comes out in the AI ecosystem for code, comes the question of migration. How do you know if Gemini 3.5 Flash is a good alternative for your business? Start by experimenting with a small group of developers over several weeks. If the results are as good as with Claude Code for a better price, consider a migration.
The key is to remain as agile and agnostic as possible regarding the technologies used. The verdict is clear for teams whose code agents run in volume on agentic workflows: the economic calculation clearly leans towards Google. For others, those who rely on complex code tasks independently, it is better to wait for Antigravity CLI to prove itself in real conditions before making a move. Anthropic will not sit idly by, and a model like Mythos could reshuffle the cards quickly.
