OpenAI launched GPT-5.6 on July 9 as a family of three models: Sol for the most demanding work, Terra as the middle tier, and Luna for fast, high-volume tasks. Three weeks later, it cut Luna’s API price by 80 percent, from $1 to $0.20 per million input tokens and from $6 to $1.20 per million output tokens.
OpenAI said the cut came from changes to the models, the software that serves them, and the agent harness that supplies context and tools. Sol also helped produce the next round of improvements. Within a human-led process, it rewrote production kernels and ran hundreds of token-generation experiments. OpenAI attributes a 20 percent reduction in serving cost to the kernel work and more than 15 percent better token efficiency to those experiments.
A frontier model was helping make its own family cheaper to run. OpenAI describes the work as a tighter feedback loop between better models and lower costs.
DeepSeek had also recently released V4, including a large model whose weights anyone could download and run under an MIT license. A capable open-weight model gives customers another option and, for teams able to operate it, a path away from an API.
I think frontier models will remain valuable, but no lab will hold an uncontested lead for long. As the model itself becomes easier to match, more of the competition moves into the tools, infrastructure, and distribution around it. For customers, that makes the ability to change models more important than betting on a permanent winner.
The lead is real, but uneven
The best closed models can still justify a premium on work where a small improvement changes the result. OpenAI’s unreleased Astra model is a recent example. OpenAI says an internal version produced ten results on long-standing open problems in mathematics and theoretical computer science, then formalized each argument in Lean.
Most model choices are less clear. Benchmarks work well when success is easy to check, such as whether a code patch passes its tests or the model gives the correct answer to a math problem. They are worse at capturing how readily a model understands the intent behind a loose prompt, whether its design or prose judgment is any good, and how much steering it needs before the output feels right. People often compress those qualities into “vibe.” The word is imprecise, but the experience is real. It should be tested on representative work rather than treated as proof on its own.
Benchmark results are less stable than they appear. The 2026 Stanford AI Index found invalid-question rates as high as 42 percent on widely used evaluations and warned that some arena rankings may partly reflect adaptation to the platform. The same report shows why the competition still matters: U.S. and Chinese models have traded the lead several times since early 2025, and the gap was 2.7 percent in March 2026.
According to DeepSeek’s model card, V4 Pro Max matched GPT-5.4 at 80.6 percent on SWE-bench Verified and reported a higher Codeforces rating, while trailing it on Terminal Bench and several reasoning measures. DeepSeek ran those comparisons, so buyers should test the models on their own work. Even so, the results put a downloadable model close enough to warrant testing for some of the same tasks as a closed frontier API.
The ranking keeps moving because the labs are improving different parts of the problem at once: training, inference, reasoning time, tools, and the harness used in evaluation. Astra may open a larger gap. Another release may narrow it. Nobody buying a model today knows how long its advantage will last.
Open weights change the negotiation
“Open” is used too loosely in AI. A release can make its trained parameters available while withholding the training data, preparation pipeline, or code needed to reproduce the model. That is why “open weights” is usually the more accurate term.
The Open Source Initiative’s AI definition reserves “open source” for systems that can be used, studied, modified, and shared, with enough information about the data and code to make those freedoms practical. DeepSeek V4’s weights and repository are under an MIT license, but the release does not provide everything that definition asks for.
In practice, open weights let a company run a model in its own environment, adapt it, or choose among hosting providers. Few companies will train a frontier model, and many will decide that operating one is not worth the hardware and staffing. They still gain an alternative they can price and test. A credible exit option affects a negotiation before anyone uses it.
Closed providers can answer by making the managed route cheaper than self-hosting. Luna at $0.20 per million input tokens is an aggressive example. OpenAI also recommends splitting work by model: Sol can resolve uncertainty and set a plan, while Luna implements well-specified changes and runs tests. The expensive model supplies judgment where it matters; the cheaper one does more of the routine execution.
The same division of labor can extend across providers. A product may use a frontier model for a difficult decision, a smaller proprietary model for volume, and an open-weight model where local control matters. Price competition does not end the value of open weights, but it can make ownership unnecessary for ordinary work.
The model is only part of the product
An agent harness connects a model to the tools and context it needs to complete a task. In Codex, the harness gives the model access to a repository and terminal, enforces sandbox and approval rules, and feeds the result of each action into the next step.
The harness can change what the same model accomplishes. OpenAI credits context management and tool routing in its GPT-5.6 efficiency work. Anthropic’s long-running application experiments found that planning, structured handoffs, and a separate evaluator helped the same underlying model complete more of the task.
OpenAI publishes the Codex CLI under Apache 2.0 while keeping the models it calls proprietary. Anthropic keeps Claude Code’s core proprietary but publishes plugins, agent SDKs, and sandboxing components. Both companies let developers extend the tooling around their agents without opening the models themselves.
That surrounding software is part of what customers pay for. Providers bundle the model with access controls, monitoring, and connections to the software a company already uses.
Those conveniences also create switching costs. A system becomes harder to move when its prompts, tools, saved context, and data access all assume one provider’s behavior. Tuning the system around one model may improve its output now while making a later change more difficult.
Teams can see the cost of switching early by keeping their tools, data, and saved context under their own control and running the same tasks through a second model.
The expensive layer is still physical
Access to weights is not access to cheap compute. DeepSeek V4 Pro has 1.6 trillion parameters, with 49 billion active for each token. Its mixture-of-experts design reduces the work done per token, but serving it still requires specialized chips, high-bandwidth memory, and engineers who can keep a distributed system efficient.
The physical layer remains concentrated. Stanford counts 5,427 data centers in the United States and says TSMC fabricates almost every leading AI chip. OpenAI says Stargate is working toward 10 gigawatts of U.S. capacity by 2029. These are signs of labs becoming infrastructure companies as well as model developers.
An open-weight model can move between operators while the chips, energy, and cloud capacity needed to run it remain controlled by a much smaller group. The software may be portable even when the machine beneath it is not.
At national scale, complete AI independence is unrealistic. A country may host a model locally and control its data while relying on foreign accelerators, fabrication, or software. Stanford’s work on the AI sovereignty paradox treats control as a set of tradeoffs rather than a binary state.
Control also depends on the provider’s data policy. Anthropic requires prompts and outputs sent to Claude Fable 5 to be retained for 30 days for safety monitoring, including in workspaces that otherwise use zero data retention. Anthropic offers ZDR arrangements for other eligible API products. Fable’s retention requirement concerns safety monitoring, not model training, but it still affects whether sensitive data can be sent to the model.
We have already seen the government intervene in access to frontier models. In June, the U.S. government ordered Anthropic to block every foreign national from Fable 5 and Mythos 5, including Anthropic employees working inside the United States. The restrictions were lifted later that month, but the models had already been taken offline worldwide. OpenAI also coordinated the initial GPT-5.6 rollout with the government and limited its preview to approved organizations. The measures were temporary, but they showed that access to a closed model can depend on more than the provider’s own policies.
Governments and companies have to decide which systems must keep running if a provider cuts access or changes its terms, which data must stay under local control, and whether a local alternative is worth the extra cost. A managed API may be sensible for most work. Public infrastructure and sensitive systems may justify a local model even when it is less convenient.
The market will stay mixed
Routine, high-volume work will move toward the cheapest model that meets the required standard. Open weights will be attractive when privacy, customization, local operation, or bargaining power justifies the extra effort. Frontier models will keep a premium where their judgment or reliability changes the outcome.
OpenAI and Anthropic are already placing their models inside the tools where teams coordinate and produce work. Anthropic’s beta Claude Tag lets a team delegate work to @Claude from a Slack thread, with access to channel context and approved tools. OpenAI’s company knowledge lets ChatGPT draw from Slack, SharePoint, Google Drive, GitHub, and other work systems.
Over time, teams embed more of their context, permissions, and recurring workflows in these products. A rival can offer a better model, but switching gets harder once daily work depends on the product around it. Providers benefit when their tools work best with their own models. Customers should decide deliberately which workflows are worth that dependence.
This forecast depends on model leads remaining temporary. If Astra or another system opens a large gap and holds it, capability may outweigh portability for longer than expected. A common, model-neutral harness standard would push the other way by making switching easier.
The model you choose will change. Use the strongest one available, but do not let it become the only place where your system’s knowledge and workflows can live.