- Claude Opus 5 (July 24) and Gemini 3.6 Flash (July 21) were released three days apart, both officially confirmed.
- Opus 5 aims the capacity of Fable 5 at around half pricewith 5-level effort control.
- Gemini 3.6 Flash costs 1.50/$7.50 per million tokens and consumes 17% fewer tokens.
- The real battle of the summer is no longer raw performance, but value for money.
In the space of three days, the two largest AI laboratories have drawn their weapons. Google launched Gemini 3.6 Flash on July 21, Anthropic responded with Claude Opus 5 on the 24th. Neither is looking for the absolute record: both are aiming for the best capacity-price ratio. For businesses, this is perhaps the most concrete news of the summer.
01 · Claude Opus 5
Opus 5: the power of Fable 5, at half the price
Anthropic positions Opus 5 as its new default model on Claude Max and the most powerful on Claude Pro. The price APIApplication Programming Interface, the programmatic access channel to a service or model. remains unchanged compared to Opus 4.8: 5 dollars per million tokens on entry, 25 on exit. The window of contextThe context window: the maximum number of tokens that a model can process in a single request (eg: 200k, 1M). climbs to 1 million tokens, the release up to 128,000 tokens, and the knowledge date reaches May 2026.
On the benchmarks advanced by Anthropic, Opus 5 doubles the score of Opus 4.8 on Frontier-Bench, surpasses Claude Fable 5 on computer usage (OSWorld 2.0) at a third of its price, and follows Fable 5 on coding (CursorBench) at half the price. On cybersecurity, however, it remains behind the specialized Mythos model.
02 · Gemini 3.6 Flash
Gemini 3.6 Flash: Google focuses everything on efficiency
On the Google side, the strategy is clear: efficiency, no flagship. Gemini 3.6 Flash costs $1.50 input and $7.50 output per million tokens, 17% cheaper output than version 3.5, while producing 17% fewer tokens for higher quality. His date of knowledge changes to March 2026.
Notable fact: Google delivered three models at once (3.6 Flash, 3.5 Flash-Lite and a 3.5 Flash Cyber variant for cybersecurity) but again rejected Gemini 3.5 Pro, deemed insufficient in coding. The next real tier is now called Gemini 4, in pre-training.
03 · The comparison
The new pricing landscape
| Model | Entry / Exit ($/M tokens) | Context | Max output |
|---|---|---|---|
| Gemini 3.6 Flash | 1.50 / 7.50 | 1M | 64k |
| Claude Opus 5 | 5 / 25 | 1M | 128k |
| Claude Opus 4.8 | 5 / 25 | 1M | , |
| Claude Fable 5 | 10/50 | 1M | , |
| GPT-5.6 Ground | 5/30 | 272k | , |
The reading is clear: for volume work (sorting, drafts, summaries), Gemini 3.6 crushes the competition on cost. For complex reasoning, Opus 5 now offers 80-90% of the capacity of the high end at half the price.
04 · Innovation
Effort control, the real novelty of Opus 5
Opus 5 lets the user choose, request by request, between cost and capacity.According to Fortune, July 24, 2026
The most commented feature is not a benchmarkA standardized test that measures the performance of models on a task (e.g. SWE-bench for coding, MMLU for general knowledge). : this is the five-level effort adjustment. In short, you decide for each request how much Opus 5 “thinks”, and therefore how much it costs. Maximum effort brings Opus 5 closer to Fable 5’s capabilities for about twice the base price, instead of five. This is a direct response to the main criticism of reasoning models: their soaring quality.
05 · Hot news
And meanwhile, at the beginning of August
The pace does not slow down. On August 1, OpenAI announced that its internal reasoning model, called Astra, had solved around ten open mathematical problems with formal proofs in Lean, without making the model public. On August 2, Alibaba followed up with Qwen 3.8 Max. The message is clear: between Chinese open models and this price war, the pressure on the margins of OpenAI and Anthropic has never been so strong.
06 · The concrete issue
What it changes for you
For an SME, an agency or a French developer, the calculation becomes unbeatable: a pipeline “Gemini 3.6 for sorting and drafts, Opus 5 for heavy reasoning” is 40 to 50% less expensive than an all-OpenAI solution, without any notable loss of quality. And as Opus 5 and Gemini 3.6 remain hosted by Western players, the sovereignty argument holds up better than with Chinese models. To compare, use our model comparator and our LLM cost calculator. See also our file on Claude Opus 4.8 and the updated Claude file.
Sources
- Anthropic, official announcement of Claude Opus 5 (July 24, 2026)
- Google DeepMind, announcement of Gemini 3.6 Flash (July 21, 2026)
- TechCrunch, “Google releases three new Gemini models, but no 3.5 Pro”
- Fortune, “Anthropic debuts Claude Opus 5 with a cost-capability toggle”
- VentureBeat, Gemini 3.6 Flash token cost reduction
Frequently asked questions
- When was Claude Opus 5 released and what does it change?
- Claude Opus 5 was released on July 24, 2026. It doubles the score of Opus 4.8 on Frontier-Bench, exceeds Fable 5 on computer usage at a third of the price, keeps the price of 5/$25 per million tokens and adds effort control at five levels.
- How much does Gemini 3.6 Flash cost?
- Gemini 3.6 Flash costs $1.50 input and $7.50 output per million tokens, 17% cheaper output than Gemini 3.5 Flash. It also consumes 17% fewer tokens for higher quality, with a context window of one million tokens.
- What is Opus 5 effort control?
- It’s a five-level setting that lets the user decide, for each query, how much the model “thinks”, and therefore how much it costs. At max effort, Opus 5 approaches Fable 5’s capabilities for about twice the base price, instead of five times.
- Is Gemini 3.5 Pro finally released?
- No. Google once again rejected Gemini 3.5 Pro, deemed insufficient in coding, while delivering three Flash models. Google’s next real tier is now called Gemini 4, currently in pre-training phase, with no date announced.
- Should you choose Opus 5 or Gemini 3.6?
- The two are complementary. Gemini 3.6 Flash is unbeatable on cost for volume tasks (sorting, drafts, summaries). Opus 5 offers 80-90% of the high-end capability at half the price for complex reasoning. Many companies combine the two.
- What happened at the beginning of August 2026 on the model side?
- On August 1, OpenAI announced that an internal model named Astra had solved around ten open mathematical problems with formal proofs in Lean, without making it public. On August 2, Alibaba launched Qwen 3.8 Max. The pace of releases remains very sustained.

