Anthropic, Meta and Thinking Machines Sign Compute Deals as Opus 4.7 and GPT-5.5 Launch
Updated: 6 days ago
Title: Anthropic, Meta and Thinking Machines Sign Compute Deals as Opus 4.7 and GPT-5.5 Launch
Date: 24 April 2026
Type: Blog
Author: SAASiQ (contact@saasiq.ai)
Word count: 1384 words
Reading time: 6 min
Published: 24-04-2026
Anthropic agreed on 20 April to spend more than $100 billion with Amazon Web Services over the next ten years, for up to 5 gigawatts of new computing capacity, and Amazon is investing $5 billion in the company now with up to $20 billion more to follow. Two weeks earlier Anthropic signed for several more gigawatts of Google's chips. Thinking Machines Lab and Meta also signed multibillion-dollar cloud contracts this month, while Anthropic, Meta and OpenAI released new models, each with its own benchmark scores.
Anthropic's agreements with Amazon and Google
The Amazon agreement runs on Trainium, the chip Amazon designs for its own data centres. Anthropic says it already uses more than a million Trainium2 chips, that more Trainium2 capacity comes online this quarter, Trainium3 later in the year, and that it will have nearly a gigawatt of the two by the end of 2026. The $5 billion comes on top of the $8 billion Amazon had already put in. Andy Jassy, Amazon's chief executive, described it as a commitment by Anthropic to run its models on Trainium for the next decade.
On 6 April Anthropic announced an agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity, the Google-designed chips that Broadcom supplies, coming online from 2027 and mostly sited in the United States. Anthropic did not give a figure, but Broadcom's regulatory filing put it at 3.5 gigawatts and said Anthropic's use of the capacity depends on its continued commercial success. That adds to a deal from October 2025 for up to one million TPUs and well over a gigawatt of capacity during 2026.
Anthropic trains and runs Claude on all three main types of chip (Trainium, TPUs and Nvidia GPUs) and says Amazon remains its primary cloud and training partner. It gave its revenue run rate as more than $30 billion, against about $9 billion at the end of 2025, and said more than 1,000 business customers now spend over $1 million a year with it, a number that had doubled in less than two months.
On 24 April, the day of this post, Bloomberg and TechCrunch reported that Google would invest up to $40 billion in Anthropic: $10 billion now at a $350 billion valuation and $30 billion more if Anthropic meets performance targets, with Google Cloud supplying 5 gigawatts of capacity over five years.
Thinking Machines signs with Google Cloud
Thinking Machines Lab, the company Mira Murati founded in February 2025 after leaving OpenAI, signed an expanded agreement with Google Cloud on 22 April. TechCrunch reported the value as in the single-digit billions of dollars. It is the lab's first contract with a cloud provider and it is not exclusive.
The agreement covers computing capacity and cloud services. Thinking Machines will use Google's A4X Max virtual machines, which run on Nvidia GB300 NVL72 racks, along with Google Kubernetes Engine, the Spanner database and Cloud Storage. Thinking Machines says early testing showed training and serving running twice as fast as on the previous generation of GPUs. Nvidia invested in Thinking Machines earlier in the year, and its seed round in 2025 raised $2 billion at a $12 billion valuation. Its first product, Tinker, a service for fine-tuning models, launched in October 2025.
Meta's first Muse model and its CoreWeave contract
Meta released Muse Spark on 8 April, the first model from Meta Superintelligence Labs. The lab is led by Alexandr Wang, who joined Meta in June 2025 when it paid $14.3 billion for a 49 per cent stake in his company, Scale AI. Muse Spark is available in the Meta AI app and on meta.ai, with a private API preview for selected users, and Meta has not released its weights.
Meta says Muse Spark took over an order of magnitude less compute to reach the same capability as Llama 4 Maverick, and that its Contemplating mode, which runs several reasoning agents in parallel, scored 58 per cent on Humanity's Last Exam and 38 per cent on FrontierScience Research. Those are Meta's own figures.
The day after the launch, CoreWeave announced a $21 billion agreement to supply Meta with AI cloud capacity through December 2032, including some of the first deployments of Nvidia's Vera Rubin platform. That is on top of the $14.2 billion contract the two signed six months earlier. Meta is also building its own sites, including Hyperion in Louisiana, which Mark Zuckerberg said in July 2025 would scale to 5 gigawatts over several years.
What the largest companies plan to spend
The deals sit inside capital budgets set earlier in the year. In January Meta guided to capital expenditure of $115 billion to $135 billion in 2026, including finance lease payments, against $72.2 billion in 2025. Alphabet said in February it expected $175 billion to $185 billion, having spent $91.4 billion on property and equipment in 2025. Amazon told investors on 5 February it expected to spend about $200 billion this year, most of it on AWS. Microsoft does not give an annual figure, but it spent $37.5 billion in the three months to December, 66 per cent more than a year earlier, and its financial year runs to the end of June.
April's benchmark scores
Anthropic announced Claude Mythos Preview on 7 April and has kept it out of general release because of how well it finds and exploits software vulnerabilities. It is available to launch partners including AWS, Apple, Broadcom, Google, Microsoft and Nvidia, and to more than 40 other organisations, under Project Glasswing. Anthropic's figures put it at 77.8 per cent on SWE-bench Pro, a test of fixing real software issues, against 53.4 per cent for Opus 4.6, and 82.0 per cent on Terminal-Bench 2.0, against 65.4 per cent.
Claude Opus 4.7 followed on 16 April, at the same price as Opus 4.6 of $5 per million input tokens and $25 per million output tokens. Anthropic reports 64.3 per cent on SWE-bench Pro, ahead of its figures for GPT-5.4 (57.7 per cent) and Gemini 3.1 Pro (54.2 per cent), and 87.6 per cent on SWE-bench Verified. It says Opus 4.7 is less broadly capable than Mythos Preview and that it reduced the model's cyber capabilities in training. Its new tokeniser can turn the same text into between 1.0 and 1.35 times as many tokens, so the bill for identical work can rise at an unchanged price per token.
OpenAI released GPT-5.5 in ChatGPT and Codex on 23 April and gives it 82.7 per cent on Terminal-Bench 2.0. Nvidia said the same day that GPT-5.5 was built and is served on its GB200 NVL72 systems, and that the two companies had brought up the first 100,000-GPU cluster of that design together.
Every one of these scores comes from the company that made the model, run on its own test setup. So Mythos Preview's 82.0 per cent and GPT-5.5's 82.7 per cent on Terminal-Bench 2.0 are two self-reported numbers rather than a head-to-head result, and Meta's figures for Muse Spark use different tests again.
Distillation and the Frontier Model Forum
Bloomberg reported on 6 April that OpenAI, Anthropic and Google had begun sharing information through the Frontier Model Forum, the industry body they founded with Microsoft in 2023, to detect adversarial distillation. That is the practice of querying a model at volume and using its answers to train a cheaper copy. It was reported as shared threat intelligence, and the three did not issue a joint statement.
On 23 February Anthropic said DeepSeek, Moonshot AI and MiniMax had created more than 24,000 fraudulent accounts and generated more than 16 million exchanges with Claude to train their own models, with MiniMax responsible for more than 13 million of them.
What buyers see
Claude is sold through Amazon Bedrock, Google Cloud's Vertex AI and Microsoft Foundry, and Anthropic says more than 100,000 customers use it on AWS. Anthropic also said a Claude Platform on AWS is coming, using the customer's existing AWS account, controls and billing.
SAASiQ's view is that the benchmark tables from April are useful for shortlisting a model, and that a pilot on the organisation's own tasks, with token counts measured, is what settles the cost.
Anthropic's new Trainium2 capacity is due online this quarter and Trainium3 later in 2026, with the TPU capacity under the Google and Broadcom agreement following from 2027.
SAASiQ - Intelligent Solutions for SaaS ©
