Capability Comes in Slices

If You Only Read One Thing

Two of today's strongest releases are defined by what users do not get. DeepSeek Prices Visual Work by compressing each image into 384 tokens; Mythos Ships Without Access because Anthropic's strongest cyber model returns vulnerability findings without an open-ended prompt surface. Capability is expanding, but labs are selling it in narrower slices rather than releasing every strength through one endpoint.

DeepSeek Prices Visual Work

DeepSeek did not merely add image input to a cheap model. It put a hard price on what an agent is allowed to see.

The new deepseek-v4-flash-vision-exp endpoint accepts screenshots, charts and document images inside the same loop that already writes code and calls tools. It works through OpenAI-compatible Chat Completions and Responses APIs as well as Anthropic-compatible Messages. The official model page lists a one-million-token context window, a 384,000-token maximum output and the same rates as text-only V4 Flash: $0.22 per million uncached input tokens and $0.66 per million output tokens off-peak, with both prices doubling during peak windows.

That removes a workflow tax. Until Friday, a visual agent using DeepSeek's API needed a second model to turn a screenshot into text, or it needed to route the whole task to a more expensive multimodal model. Native vision lets one model inspect an error dialog, reason about the repository and call the next tool without translating between providers.

The mechanism is bounded observation. Think of it as a postage stamp for visual context: DeepSeek resizes each large image toward 800 by 800 pixels and caps it at 384 input tokens. At off-peak rates, that ceiling costs about $0.000084 per image before text, output and retries. The low bill comes from discarding resolution, not from making full-fidelity vision free.

That trade creates a new model split. Cheap vision can classify screenshots, triage ordinary documents and read uncomplicated charts. Fine interface controls, tiny labels and dense tables can disappear during resizing. The likely architecture is not one visual model winning every task. It is a low-cost observer handling the first pass while detail-sensitive cases escalate to a higher-resolution model.

The strongest counterargument is that DeepSeek has not proved even the first-pass role. The endpoint is explicitly experimental, its architecture is undisclosed and its launch scores are vendor-run. DeepSeek says visual-agent performance approaches Claude Opus 4.8, but the release mixes text-only and visual evaluations under its own harness. A current specification audit confirms the price and resolution boundary, not equivalent task quality.

The payoff is a broader model choice for browser and document agents, with fidelity now visible as the constraint. The concrete test is September 30: promotion from -exp to a stable model ID at the same image ceiling and no more than a twofold price increase would show that bounded observation is a product tier rather than a launch experiment.

Mythos Ships Without Access

Anthropic has made its strongest cyber model more available by withholding the thing model releases usually sell: direct access to the model. Enterprise customers can now run Mythos 5 inside Claude Security, but they receive vulnerability findings and suggested patches rather than an open-ended Mythos conversation.

The August 21 release changes a boundary Anthropic drew in April. Project Glasswing gave Mythos-class access to a small set of organizations protecting critical software. The broadly available Fable 5 model then added safeguards that block risky cyber work. Mythos itself remained behind a narrow admission process.

Claude Security turns that admission problem into a product shape. An Enterprise administrator enables the public beta, a user selects a repository, and Mythos scans the code for vulnerabilities. Each result includes a standard weakness category, confidence and severity ratings, and a suggested fix. The scan is billed as ordinary token usage under the existing Enterprise plan.

The crucial limitation comes after the result. A suggested patch opens in Claude Code on the web and runs on models the organization already has. Mythos does not follow the user into that interactive session. Anthropic exposes the security artifact while keeping the more dangerous general capability behind the service.

This is outcome-gated access. Think of it as hiring a specialist through a service window rather than handing over the specialist's full toolkit. The customer submits one bounded job and receives one bounded artifact. Anthropic controls the surrounding workflow, which makes offensive repurposing harder and lets the lab distribute more capability than an unrestricted API would permit.

Last Saturday's GLM analysis described a temporary two-week delay before raw weights. Anthropic is taking the opposite route. It can widen defensive use now without promising that the underlying model will ever become generally promptable. The unit of distribution is no longer always a checkpoint or endpoint. It can be a specific result produced behind the vendor's boundary.

The strongest counterargument is that Claude Security already existed. Anthropic has changed the model behind a scanner, not created a new workflow. The company also provides no independent precision, recall or false-positive comparison against the previous system. A stronger hidden model is not automatically a better security product if reviewers cannot distinguish useful findings from plausible noise.

The structural payoff is a third release tier between closed API access and no access at all. Labs can sell high-risk capability through task-specific products while keeping general control. That makes the interface part of the safety system and the business model at once.

The concrete test is partner distribution. By September 30, at least one external security product should ship a Mythos-backed scan and report validated findings or reviewer acceptance against its prior system. If access remains confined to Anthropic's own beta with no comparative outcomes, the release widened the funnel without proving the capability travels.

The Contrarian Take

Everyone says: Better models expand access: cheaper vision reaches more agents, while stronger cyber models put frontier capability in more hands.

Here's why that's wrong (or at least incomplete): Both releases expand use by narrowing the product. DeepSeek compresses every image into at most 384 tokens, trading detail for price. Anthropic exposes Mythos through a scanner that returns defensive artifacts without giving users a general Mythos prompt surface. The new competition is not simply who has the strongest model. It is who can package the right slice of capability at a usable price and risk boundary.

Under the Radar

Model routing can destroy the saving it seeks — Copilot's production traces show cache reuse falling from roughly 90% within a turn to 8% after a model switch. A router can choose a cheaper next model and still pay to rebuild a huge prompt prefix, so per-token price and per-session cost can move in opposite directions. (Study)
DeepSeek lets tools return images directly — Its Responses API accepts image content inside tool outputs, not only in the initial user message. That makes visual state part of an iterative agent loop, while the 384-token ceiling keeps every new observation bounded. (Guide)

Quick Takes

Terminal-Bench made the result movable without making it vague. Version 3.0 adds semantic versions, continuous task updates and result migration, so trials can be re-graded instead of discarded when a task changes. Its early results also separate task completion from tokens and cost, which makes equal scores less equal. (Source)
Agent tools are spreading faster than their ownership model. A scan of 33,563 published MCP server builds found nearly half raised at least one security finding, about one in eight exposed an irreversible first-call action and 84% of identifiable maintainers were solo publishers. A shared tool protocol has become a software supply chain before it acquired institutional review. (Source)
Enterprise agent adoption has an activation denominator. Salesforce telemetry from 400 Agentforce customers says active agents per organization nearly tripled over five quarters. That is credible evidence of production spread inside one platform, but an activated agent is not a completed task or a dollar of return. Usage has outrun a portable outcome measure. (Source)

The Thread

Capability is being sold in slices. DeepSeek packages vision as a low-cost observation that discards fine detail. Anthropic packages Mythos as a security result without a general-purpose model session. Terminal-Bench adds a third cut by separating similar task scores through token use and cost. Model releases once meant a new endpoint with a broad capability envelope. The emerging product is a bounded job, with fidelity, price and risk chosen before the model runs.

Predictions

New predictions:

I predict: By September 15, an independent visual-agent evaluation will show DeepSeek V4 Flash Vision Exp trailing Claude Opus 4.8 by at least 10 percentage points on detail-sensitive screenshot or computer-use tasks, despite a much lower token bill. (Confidence: medium; Check by: 2026-09-15)

Issue date: August 22, 2026 · Generated: 04:09 AM ET