Generated by Codex with GPT-5

Techmeme surfaced this June 5, 2026 story in its xAI and Claude cluster, and the original article is The Information’s How xAI Went From Chasing Anthropic to Powering It.

The story is interesting because it compresses several frontier-AI tensions into one company drama. The Information reports that xAI used Claude outputs while trying to catch Anthropic in coding models, including a multi-month distillation effort, personal-account workarounds after access was cut off, and access through Blackbox AI for benchmarking and other work. At the same time, xAI and the broader SpaceX orbit have been moving into compute partnerships with Anthropic and Cursor, raising the question of whether the strategic center is shifting from “build the best model” to “control scarce infrastructure and distribution.”

That makes the article more than another report about Musk-company turbulence. It shows how hard it is becoming to separate model competition, developer-tool strategy, training-data provenance, cloud economics, and terms-of-service enforcement. In a market where the best coding models can directly improve the next generation of coding products, access to a rival model is no longer just a convenience. It can become part of the production process.

The Coding-Model Gap

xAI entered the coding model market late. By the time its first coding-focused model appeared, Anthropic and OpenAI had already established themselves as the reference vendors for AI-assisted software work. Claude in particular had become a high-value target because its coding behavior is useful not only to end users, but also to rival labs trying to learn what strong agentic coding looks like.

The Information’s reporting says xAI kept using Claude outputs even after Anthropic began cutting off rival access. The sharpest allegation is that xAI used Claude responses in a multi-month distillation project to train its own coding model. Distillation is not the same thing as ordinary benchmarking. Benchmarking asks, “How good is the other model?” Distillation asks, in effect, “Can our model learn from the other model’s answers?”

That distinction matters because API terms, research norms, and business expectations are all being stress-tested at once. Frontier labs sell access to powerful models because usage drives revenue and adoption. But broad access also makes it easier for competitors to extract behavior, compare outputs at scale, and train downstream systems. The better a model is, the more attractive its outputs become as training material.

The report also describes engineers using personal Claude accounts after xAI’s company access was cut off, and later using Blackbox AI, an intermediary service that routes access to many models. That is the practical enforcement problem in miniature. A lab can terminate an account. It cannot easily prevent every employee, vendor, router, reseller, evaluation harness, and proxy from touching its model. As more tools aggregate access to multiple model providers, the boundary around a proprietary model gets harder to police.

Compute Changes The Leverage

The strangest part of the story is that xAI was reportedly chasing Anthropic’s coding capability while also becoming useful to Anthropic as an infrastructure provider. In May, xAI announced a compute deal with Anthropic reportedly worth \$1.25 billion per month for a major slice of xAI’s data-center capacity. The same company that wanted to learn from Claude was now leasing critical capacity to Claude’s maker.

That is the strategic inversion. If frontier model quality is difficult to sustain, compute can still be valuable. SpaceX and xAI have spent heavily on data-center capacity, and scarce capacity can be monetized even when model development is uneven. The Information frames this as an open question around SpaceX’s AI effort: is the long-term business the models themselves, or the infrastructure that other leading AI companies need?

The Cursor relationship points in the same direction. xAI agreed to give Cursor access to its infrastructure to train Composer, and SpaceX reportedly has an option to buy Cursor for \$60 billion or pay a \$10 billion breakup fee. Cursor brings a dominant developer workflow and post-training expertise. xAI brings infrastructure. The package suggests a more modular AI stack: one party has users and product taste, another has compute, another has model research depth.

That modularity can be powerful, but it also exposes xAI’s weaknesses. The Information reports staff churn, leadership turnover in the coding project, an accidental deletion of important training data, a smaller pretraining team than before, and talks with Mistral that did not turn into a formal partnership. Those details make the compute strategy look less like an optional second business and more like a hedge against the difficulty of building frontier models under organizational strain.

Access Is Becoming A Control Plane

The article also shows that model access is becoming a competitive control plane. Anthropic cutting off OpenAI and xAI was not just a customer-management decision. It was a way to protect model behavior from rivals that could use Claude to build or tune competing products.

That may become standard. Frontier labs are likely to tighten API terms, monitor suspicious usage patterns, watermark or fingerprint outputs where possible, and limit bulk access that looks like training-data collection. But enforcement will remain incomplete. The most useful customers often run large volumes of highly technical prompts, and that can look similar to adversarial benchmarking or distillation. A sophisticated rival can spread work across accounts, contractors, subsidiaries, model routers, or enterprise workflows.

The Blackbox AI detail is important for that reason. Model aggregators create convenience for developers and businesses, but they also complicate provenance. When a request passes through an intermediary, the upstream lab may have less context about who is using the model and why. That does not make intermediaries illegitimate. It does make them part of the trust boundary in a way that the AI industry has not fully settled.

For developers, this can feel abstract until a tool they rely on changes access overnight. If coding agents are built on a shifting mix of first-party models, third-party models, distillation, routers, and private infrastructure deals, then product reliability depends on commercial permissions as much as technical performance. A model can get worse, a vendor can close access, a router can lose a provider, or a partner can decide the data-sharing risk is too high.

Why Coding Agents Make This Sharper

Coding is a particularly sensitive frontier because it is measurable, valuable, and self-reinforcing. A strong coding model can generate patches, tests, evals, scaffolding, refactors, and training data. It can also help build the tooling used to improve the next model. That makes a rival’s coding model both a product competitor and a source of process knowledge.

If one lab can cheaply query another lab’s coding model and train on the answers, then the gap between first mover and follower can narrow. That is why the industry is becoming more aggressive about access restrictions. The expensive part of frontier work is not only the final weights. It is the accumulated taste in data, post-training, tool use, evaluation design, and failure analysis. Outputs leak some of that taste.

But copying outputs is not a complete shortcut. A model trained from another model’s responses may inherit useful patterns, but it does not automatically inherit the original lab’s data pipeline, internal eval culture, infrastructure discipline, product feedback loop, or research judgment. The Information’s account of xAI’s coding effort suggests that the organization around the model matters as much as the borrowed signal. Training on good answers is easier than building the machine that consistently produces them.

That is the deeper lesson for AI coding companies. The moat is not just model quality on a leaderboard. It is the whole system: access to compute, access to users, permission to use training signals, reliable infrastructure, product integration, security posture, eval discipline, and the organizational ability to keep improving without breaking trust.

Takeaway

Techmeme was right to surface this because it is a concrete example of the AI industry’s next competitive phase. The frontier is no longer only a contest over who can train the smartest model. It is also a contest over who can access whose outputs, who can enforce terms, who controls compute, who owns developer workflows, and who can turn all of that into a durable product.

The xAI story is messy, but the mess is informative. A company can be a rival, a customer, an infrastructure supplier, a distribution partner, and a suspected terms-of-service problem at the same time. That is what happens when the inputs to AI progress are scarce, the outputs are valuable, and the same few companies sit on both sides of the market.

The practical thing to watch is whether leading labs start treating model outputs more like regulated supply chains. Provenance, permitted use, routing transparency, and auditability may become as important to AI competition as benchmark scores. If coding models can help build coding models, then access policy is no longer legal boilerplate. It is part of the frontier.