Kimi K3 drives prices down, Anthropic plays the stock-market card

Ten days after the “Super Thursday” of July 9, when OpenAI and xAI launched GPT-5.6 and Grok 4.5 on the same day, the picture is getting sharper and more complicated. A Chinese model muscles into the world top 4 at a slashed price, xAI finally publishes the price of its flagship, and Anthropic chooses to answer the competitive pressure not by cutting its prices, but by quietly filing for a stock-market listing. Two parallel races are taking shape: the race on prices, which keeps falling, and the race on valuations, which keeps climbing.

Kimi K3: China muscles into the world top 4

Moonshot AI put Kimi K3 online on July 16, an open-weight model of 2.8 trillion parameters (MoE, only 16 of 896 experts activated per token), with a context of one million tokens and an always-on reasoning mode. On the independent benchmarks published in its wake, it ranks as the 4th frontier model in the world, behind Claude Fable 5 and GPT-5.6 Sol but ahead of Claude Opus 4.8, while billed at $3 / $15 per million tokens (input/output), with a cache-hit rate of $0.30. It is the first open-weight model to climb this high in the ranking, and it is available both through the Moonshot API and through aggregators such as OpenRouter, which opens the way to sovereign hosting for organisations that prefer not to depend on a single access point.

Grok 4.5 finally has a price, but not yet for Europe

On July 9, xAI had launched Grok 4.5 with no API price and no independent benchmarks. That is now settled: $2 / $6 per million tokens, more than 60% cheaper than Claude Opus 4.8 or GPT-5.5, for a model that Elon Musk calls “Opus-class, but faster and cheaper”. The model was trained in part on real Cursor developer sessions, with a clear positioning on code and agentic tasks. A point of caution for an organisation based in Switzerland: at launch, Grok 4.5 remained unavailable in the xAI API and products in the EU zone. An extension in mid-July was announced but has not been confirmed to date, so check before any routing decision.

Comparison table: API prices as of July 19, 2026

Model	Input ($/M tokens)	Output ($/M tokens)	Note
Claude Fable 5 (Anthropic)	10.00	50.00	credits add-on
GPT-5.6 Sol (OpenAI)	5.00	30.00	launched July 9
Claude Opus 4.8	5.00	25.00
Kimi K3 (Moonshot, China, open-weight)	3.00	15.00	launched July 16; cache hit 0.30
GPT-5.6 Terra	2.50	15.00	launched July 9
Gemini 3.1 Pro (Google)	2.00	12.00	doubles beyond 200k context
Claude Sonnet 5	2.00	10.00	launch price until August 31, then 3/15
Grok 4.5 (xAI)	2.00	6.00	launched July 8; price since published; EU restricted
Gemini 3.5 Flash	1.50	9.00
Grok 4.3 / 4.20 (xAI)	1.25	2.50	1M context
GPT-5.6 Luna	1.00	6.00	launched July 9
Mistral Large 3	0.50	1.50
DeepSeek V4 Pro	0.435	0.87	after the May cut
DeepSeek V4 Flash	0.14	0.28	1M context; cache up to −99%

Prices recorded on July 19, 2026 from official pages (Anthropic, OpenAI) and specialised aggregators for models without a directly accessible pricing page.

Anthropic answers with the stock market, not with price

Rather than follow the general decline, Anthropic has chosen another playing field. On June 1, the company filed a confidential IPO registration (S-1) with the SEC, following a $65 billion funding round on May 28 that valued it at $965 billion, overtaking OpenAI for the first time. Its annualised revenue is said to have passed $47 billion, driven by the mass adoption of Claude Code in enterprises. Several outlets report that, instead of lowering its rates against Grok 4.5 and Kimi K3, Anthropic would rather have relaxed the usage caps on its subscriptions to retain developers. This is not officially confirmed at this stage and should be taken with caution.

On the other side, OpenAI filed its own confidential paperwork around May 22, on the basis of a $110 billion round closed in late February at a $730 billion valuation ($840 billion post-money), and is said to be aiming for a listing at more than $1 trillion in September. The same two companies are therefore running, in parallel, a race to the bottom on token prices and a race to the top on their own valuations. The paradox is only apparent: it is volume of usage, not margin per token, that has to convince the markets.

Compute, the blind spot of falling prices

Token prices are falling, but the infrastructure behind them costs more. The data-centre GPU market reaches $48.4 billion in 2026, and AWS has raised the price of its premium EC2 Capacity Block instances by 15% (the p5e goes from $43.26 to $49.75 per hour in US West). Renting an H100 now runs between $2.01 and $11.06 per hour depending on the provider, and owning or colocating your GPUs remains 40 to 60% cheaper than the public cloud beyond 65-70% sustained utilisation. A sign of the times on the financial markets: Apple overtook Nvidia as the world’s most valuable company on July 17, the first time the stock-market summit of AI is not occupied by a chip maker.

What this means for a Swiss organisation

Three concrete points of attention. First, open-weight becomes a credible option again: with Kimi K3 in the world top 4, sovereign self-hosting is no longer a performance compromise, only a governance and compliance choice. Second, check regional availability before choosing: Grok 4.5 shows that a published price does not guarantee immediate access from Switzerland or the EU. Finally, pricing commitments from vendors racing to an IPO can change fast: a company aiming at a trillion-dollar valuation has different incentives from one that simply sells tokens, something to watch over the coming months.