The week AI companies stopped being model companies

On one side of this week's AI market: an 80% price cut. On the other: $15.83 billion in quarterly investment. The figures belong to different vendors and different businesses, but the distance between them explains week 32 better than another model ranking.

Perplexity made some agent work dramatically cheaper. At the same time, xAI's broad AI segment spent heavily on compute and described its next move toward chip manufacturing. Google started packaging model routing across providers. Anthropic moved more work into persistent cloud environments. Manus joined conversational work to audio production. Mistral released an open specialist model for moderation.

The week's thesis is deliberately blunt: AI companies are starting to stop being model companies. They are becoming operating companies that sell the route from raw compute to the place where work gets finished.

The model became the cheapest part of the offer

According to this week's research, Perplexity's Agent API cut Luna pricing by 80% and Terra by 20%. The same vendor launched Computer for Builders, where GitHub, Vercel, Datadog, Supabase, Stripe, and Slack can sit inside one chain from code to operations and customer communication.

Those are two different businesses. The model call can fall in price while value moves to orchestration, connectors, scheduled runs, and the place where several systems meet. This does not prove that Computer completes the entire chain without friction. It shows that Perplexity wants to own more of the work surface than the answer itself.

xAI shows the other end of the same market. SpaceXAI's reported AI segment produced $2.561 billion in second-quarter revenue and invested $15.83 billion. The segment is broader than Grok, so those figures are not the model's own profit and loss statement. They still reveal the company xAI is trying to build: models, data centers, networks, and planned chip manufacturing inside one structure. The announced first phase of Terafab was put at more than $16.8 billion. That is a plan, not finished chips or lower API prices.

Cheaper intelligence at the edge may rest on an increasingly expensive industry behind the screen.

The product starts before the model and ends after it

Google offered a clear example with model routing in API Gateway. In Public Preview, one interface can send text requests to Gemini, Claude, or open GPT models hosted through Vertex AI Model Garden. The router does not magically choose the best or cheapest model. It follows configured rules and model names. The product is the switchboard: authentication, quotas, traffic visibility, and the ability to change the destination behind one client contract.

Google also backed Agent Plugins as a portable package format and released WeatherNext Cyclones as a specialist research system for weather forecasting. One effort tries to standardize how agent capabilities travel between tools. The other shows that valuable AI does not always need to be a general chat model.

Manus and Mistral filled in different parts of the picture. The new Manus connection to ElevenLabs makes speech generation, transcription, and voice apps part of one conversation. Mistral's Shieldstral 1.0 is an open specialist model that evaluates text and images against a policy written in ordinary language. Neither update is merely “a smarter model.” One packages a production process. The other provides a component that can run in an organization's own technical environment.

OpenAI showed why a model name is not enough

OpenAI supplied the week's clearest warning for anyone comparing models by the label in a menu. The August build of GPT-5.6 Sol rolled out in ChatGPT for Plus and Pro, while Work and Codex remained on July builds. The API change concerned faster handling for very long contexts. The same family name could therefore hide different versions, product surfaces, and operating terms.

This is more than version administration. When a vendor tunes the same model family differently for consumer chat, work, coding, and API use, the product context becomes part of what a buyer is purchasing. A test in ChatGPT does not automatically describe how the same name behaves in Codex or a custom workflow.

OpenAI was building in several directions at once: voice architecture, education packages, content-provenance checks, agent SDKs, and a public dataset about usage patterns. The model remains central, but it is no longer the whole business.

The competition is about who operates the work

Anthropic expanded Cowork to more plans and made the distinction between local and cloud execution hard to ignore. A cloud session can continue after the user closes the computer, use connectors, and run inside an Anthropic-managed environment. Anthropic also introduced controls that scan skills and plugins and let an enterprise block some prompts before inference.

That sounds like enterprise governance, but the competitive consequence is wider. The vendor operating the runtime has a deeper role than a vendor returning an answer. It owns part of the waiting time, file flow, failure handling, and evidence when something goes wrong.

This week's incidents show the cost of that role. Anthropic had a multi-model degradation lasting just over seven hours. Perplexity reported a Computer sandbox issue lasting three and a half hours. Once a service is sold as a work layer, an outage is no longer only an empty chat box. It can leave a workflow half-complete across several systems.

The counterargument: a model leap can still reset the market

The thesis should not be pushed too far. A major improvement in reasoning, cost, or multimodal capability can still move the market in days. OpenAI, for example, reported clear factual-reliability gains for its August GPT-5.6 builds, although the figures came from the vendor's own evaluation and cannot be transferred directly to every workload.

The model has not become unimportant. The unit of comparison has expanded. The old question was often “which model is best?” The more useful question now is which combination of model, runtime, connectors, commercial terms, and evidence fits the work.

Listen for the business around the model

The podcast covers Claude, Gemini and DeepMind, Grok, Manus, Mistral, OpenAI, and Perplexity one by one. Listen for three things:

• Where is the vendor's real investment? The model, infrastructure, distribution, or workflow?
• What is the product after the model answers? A file, a deployed service, a decision, a conversation, or simply more text?
• Which part can actually be switched? The model name may be replaceable even when the surrounding work surface is not.

If your organization notices that a model choice is turning into a choice of an entire technical environment, compare those layers separately. Hammer Automation can help with that comparison without starting from a vendor's product catalog.

This episode is an AI-generated masterclass built from Hammer's daily deep research into AI-provider updates and features, processed with NotebookLM. The source window is the completed week of August 3–9, 2026; the provider research used in the episode was published August 3–7.

Source basis: Hammer's AI-generated masterclass research, based on daily deep research into each provider's updates and features.