What happened on 19 May 2026
While the press focused on Samsung's smart glasses and the Gemini Spark agent, Google quietly dropped the real technical bombshell of its keynote: Gemini 3.5 Flash. As we documented in our full article on Google I/O 2026, the announcement was almost buried under the hardware demos. That is a shame, because this model is likely to change the economics for millions of developers and businesses.
The detail worth pausing on: a "Flash" model (so small and cheap) beats Google's previous "Pro" on coding and agentic AI benchmarks. And it competes seriously with the competition's high-end models. At 1.50 dollars per million input tokens, it is probably the best price-performance ratio ever launched on the market.
Sundar Pichai also confirmed that Gemini 3.5 Pro will arrive in June 2026. The keynote audience applauded audibly, a sign that it is probably the most anticipated model of the year. In the meantime, Flash is already available. Here is the full technical breakdown.
The raw numbers that speak for themselves
Before the analysis, let us set out the benchmarks officially published by Google on launch day. Here are the scores on the evaluations most relevant to developers' real work.
| Benchmark | Gemini 3.5 Flash | Gemini 3.1 Pro | GPT-5.5 | Opus 4.7 |
|---|---|---|---|---|
| MCP Atlas (agentic) | 83.6% | 78.2% | 75.3% | ~78% |
| Terminal-Bench 2.1 (coding) | 76.2% | 70.3% | 78.2% | ~75% |
| SWE-Bench Pro (real-world coding) | ~58% | ~55% | ~61% | 64.3% |
| ARC-AGI-2 (reasoning) | 72.1% | ~75% | 84.6% | ~80% |
| Humanity's Last Exam | ~35% | ~38% | ~42% | 46.9% |
| MMMU-Pro (multimodal) | 83.6% | ~82% | 81.2% | ~80% |
| MRCR v2 (long context 128k) | 77.3% | 84.9% | 94.8% | ~88% |
| Speed (tokens/second) | 289 t/s | ~95 t/s | ~75 t/s | ~70 t/s |
Across 5 benchmarks from Google's official table (MCP Atlas, Toolathlon, Finance Agent v2, CharXiv Reasoning, MMMU-Pro), Gemini 3.5 Flash beats both Claude Opus 4.7 AND GPT-5.5. Yet Flash is supposed to be the "small" model in Google's line-up. That means one thing: the boundary between Flash and Pro is now far blurrier than it was two years ago. Marketing says "Flash = fast and cheap", reality says "Flash = as good as premium competitors on several critical tasks".
Price: Google's nuclear weapon
The whole appeal of Gemini 3.5 Flash crystallises around its price. Here is the raw comparison:
| Model | Input / M tokens | Output / M tokens |
|---|---|---|
| Gemini 3.5 Flash | $1.50 | $9 |
| Gemini 3.1 Pro | $2 | $12 |
| GPT-5.5 | $5 | $30 |
| Claude Sonnet 4.6 | $3 | $15 |
| Claude Opus 4.7 | $15 | $75 |
Let's do the concrete maths. For an agentic workflow that consumes 10 million input tokens and generates 2 million output tokens:
- Gemini 3.5 Flash: 15 + 18 = $33
- Claude Sonnet 4.6: 30 + 30 = $60 (1.8x more expensive)
- GPT-5.5: 50 + 60 = $110 (3.3x more expensive)
- Claude Opus 4.7: 150 + 150 = $300 (9.1x more expensive)
At industrial volumes, these gaps quickly become dizzying. A company consuming 1 billion tokens per month saves around $7,700 by switching from Opus 4.7 to Gemini 3.5 Flash, and around $2,700 by switching from GPT-5.5 to Flash. Multiplied over 12 months, these are five- or six-figure decisions at stake.
Who wins at what: the honest table
Beyond the raw figures, here is who is better at what in May 2026. This is the honest analysis that few comparisons take the trouble to make.
Gemini 3.5 Flash clearly dominates on:
- Multi-tool agentic AI (MCP Atlas, Toolathlon, Finance Agent v2). If you are building agents that orchestrate several APIs, Flash is probably the best choice.
- Multimodality (MMMU-Pro). Understanding images, videos, complex PDFs: Flash beats all the others.
- Raw speed. 289 tokens per second is 4 times faster than Opus 4.7 and GPT-5.5. For responsive user interfaces or industrial workflows, that is decisive.
- The price-performance ratio. No other model comes close.
GPT-5.5 keeps the edge on:
- Complex reasoning (ARC-AGI-2: 84.6%). For difficult mathematical problems and deep logical analysis, GPT-5.5 remains ahead.
- Ultra-reliable long context (MRCR v2 at 128k: 94.8%). When you load an entire book into the model, GPT-5.5 has the best retention.
- Terminal-Bench 2.0 (systems coding) at 82.7%. For complex command-line operations, OpenAI remains best placed.
Claude Opus 4.7 dominates on:
- SWE-Bench Pro (64.3%): the reference benchmark for real coding on existing repositories. For an existing Python or TypeScript project, Opus 4.7 remains the benchmark.
- Humanity's Last Exam (46.9%): the industry's hardest benchmark, testing deep knowledge and scientific creativity.
- Long, complex tasks requiring consistency and creativity (long-form writing, multi-step reasoning over several hours).
Claude Sonnet 4.6 remains relevant for:
- Everyday workflows where you don't need Opus but want Claude's quality (instruction following, natural tone)
- Enterprise deployments where stability and predictability matter more than extreme benchmarks
The benchmarks Google published on 19 May are a snapshot of Google itself. It chooses the tests, the conditions, the configurations. Several independent experts (at Hugging Face, at Stanford, on Aider, on LMArena) are currently re-testing all these models under real-world conditions. Early feedback suggests that the gaps between Flash, GPT-5.5 and Opus 4.7 are sometimes narrower than Google's official table suggests. Never take any vendor benchmark literally, ever.
How to choose in practice in May 2026
To help you decide, here is a practical guide based on your use case.
If you are building an AI agent that orchestrates tools. Gemini 3.5 Flash is probably the best choice, especially if you work with the MCP protocol (which we explored in our article on MCP and the disruption of freelancers). The MCP Atlas score and the unbeatable price are decisive.
If you write or modify complex code on an existing project. Claude Opus 4.7 remains the reference (64.3% on SWE-Bench Pro). It is expensive ($15/$75 per million tokens), but on critical production code, the quality gap often justifies the extra cost. As we analysed in our article on Cursor and the winning AI startups, it is in fact mainly Claude that powers Cursor.
If you do difficult scientific or mathematical reasoning. GPT-5.5 keeps the edge on ARC-AGI-2 and Humanity's Last Exam. For very open-ended problems, OpenAI remains ahead.
If you process long text (books, contracts, massive codebase). GPT-5.5 on MRCR v2 at 94.8% is unbeatable for fidelity in 128k+ contexts. Claude Opus 4.7 in 1M context beta is the high-end alternative.
If you are building a cost-optimised consumer application. Gemini 3.5 Flash, without hesitation. The price-performance ratio has no equivalent. The speed of 289 tokens/second enables ultra-responsive interfaces.
If you are an independent developer who wants to experiment. Gemini 3.5 Flash is free in the Gemini app, AI Mode in Google Search, and Google AI Studio. No credit card required to test it seriously. It is probably the best entry point for learning modern AI capabilities in 2026.
What this tells us about the model race
Beyond the technical comparison, this launch tells us three things about the state of the AI war in May 2026.
First, the commodification of models is accelerating. When a "Flash" model beats the competition's "Pro" models on certain benchmarks at a third of the price, it is a sign that the industry is reaching economic maturity. As we analysed in our article on the AI bubble, this commodification is probably healthy in the long term but painful for those who bet on premium margins.
Second, Google has a structural advantage. The TPU 8i announced at the same keynote (which we covered in detail in our Google I/O article) give Google a marginal inference cost far lower than OpenAI and Anthropic, which depend on Nvidia GPUs. That difference shows up directly in the pricing: Flash at $1.50 vs GPT-5.5 at $5, that is partly the hardware.
Third, diversification remains the best strategy. No single model dominates across all benchmarks. Cursor uses Claude for coding and OpenAI for certain tasks. Perplexity uses up to 19 different models depending on the query. The right practice in 2026 is no longer "choose a model", it is "build an orchestration layer that routes to the right model depending on the task". That is exactly what the AI startups that outperform have understood.
And what about Gemini 3.5 Pro?
If Flash already does all that, what will Pro bring in June? According to leaks and official statements, Gemini 3.5 Pro should:
- Significantly improve scores on the benchmarks where Flash lags behind (ARC-AGI-2, SWE-Bench Pro, Humanity's Last Exam)
- Push the context window beyond a stable million tokens
- Probably cost in the region of 3-5 dollars per million input tokens, still cheaper than GPT-5.5 and well below Opus 4.7
If these estimates are confirmed, Gemini 3.5 Pro could be the first model to dominate on both quality and price at the same time. Anthropic and OpenAI would then be under major pressure to respond. Probably with a unilateral price cut, as we have already seen with GPT-5o vs GPT-4 Turbo.
Our practical verdict for Alyriaz
To sum up in three lines what we would recommend in May 2026:
- For 80% of everyday uses (chat, simple code, search, content): Gemini 3.5 Flash. Speed, price, quality, free access in the Gemini app. It has become the best starting point.
- For professional coding on serious projects: Claude Opus 4.7. The cost is high but the reliability on SWE-Bench Pro and the natural tone justify the premium price.
- For extreme reasoning or long contexts: GPT-5.5. Still the leader on ARC-AGI-2 and MRCR v2.
And above all: test. No comparison on paper replaces putting it into practice on your use case. Benchmarks are indications, not truths. The right model for you depends on the exact nature of your work, your tolerance for error, and your actual budget.
One thing is certain: with Gemini 3.5 Flash, Google has just rebalanced the economics of AI for millions of developers. The question is no longer "can I afford AI in my product". It is "how do I structure my orchestration to maximise quality per dollar spent". And that is probably the best question one can ask in May 2026.