DeepSeek Cuts Output Costs While One Prompt Fabricates Google Earth Scenes

01DeepSeek’s V4 Flash beta takes one model-name change; output costs $0.28 per million tokens

DeepSeek put V4 Flash 0731 into public beta on July 31 without changing its API calling method. An existing integration can reach it by setting the model name to deepseek-v4-flash. That reduces a first trial to a configuration edit, rather than a new interface integration.

The route extends the setup DeepSeek introduced on April 24. Both V4 Pro and V4 Flash already worked through OpenAI’s ChatCompletions interface and Anthropic’s interface. DeepSeek kept the base URL unchanged, allowing teams to test the update through their existing client structure.

The narrow switch also defines the release. DeepSeek says 0731 retains the preview model’s architecture and size; the company only re-post-trained it. The update applies solely to the V4 Flash API, limiting the variables developers must isolate when comparing versions.

DeepSeek reported an 82.7 score on Terminal Bench 2.1 and 76.7 on Cybergym. DeepSWE came in at 54.4, while the verified Toolathlon result reached 70.3. These scores cover coding, tool use and agent tasks without requiring developers to infer performance from a general-purpose benchmark alone.

The testing setup matters. For public code-agent benchmarks, DeepSeek used its forthcoming Harness in minimal mode, with maximum effort, top-p 0.95 and temperature 1.0. The company has not yet released that harness.

Two other reported results came from internal sets: 68.7 on DSBench-FullStack and 59.6 on DSBench-Hard. DeepSeek labels both as internal, unlike the named public benchmark sets available to outside evaluators.

Artificial Analysis subjected the reasoning model at maximum effort to a separate intelligence and price assessment. V4 Flash scored 50 on its Intelligence Index, double the 25 median for comparable models. It generated 210 million tokens during testing, versus a 100 million median.

That output volume complicates the sticker price. Artificial Analysis listed input at $0.14 per million tokens, against a $0.43 comparison median. Output cost $0.28 per million, versus $1.20. The site restricts open-weight comparisons to models within the same size class.

Running the full Intelligence Index cost $72.02. Developers now have public API access, disclosed test settings and an external cost baseline for measuring their own workloads.

02Chrome fixed more bugs in June than in two years; Anthropic is investigating three evaluation incidents

Google says AI helped Chrome fix more security bugs in June than during the previous two years combined. The gain did not come from vulnerability discovery alone. Google deployed models across a pipeline that includes finding, triaging, fixing, releasing, and applying patches.

That distinction sets the operational stakes. Finding hundreds of bugs creates little protection if engineers cannot classify them, produce safe patches, and ship updates before attackers build exploits. Chrome’s process ends only when users restart the browser and apply the release.

Google began using large language models to expand fuzzing coverage and performance in 2023. The next year, it worked with Project Zero on Naptime, which equipped models with specialized vulnerability-research tools. The company now says it is deploying AI at scale to find and fix hundreds of security bugs.

Those systems increase defensive capacity. They also increase the number of model-driven security operations that teams must authorize, isolate, and monitor.

Anthropic has published a separate item titled “Investigating three real-world incidents in our cybersecurity evaluations.” Its available summary provides no incident details, causes, or outcomes. The title alone establishes a sharp boundary: these were evaluation incidents, not disclosed product vulnerabilities or routine remediation failures.

The categories demand different controls. Vulnerability discovery asks whether a model can identify exploitable code. Remediation asks whether teams can validate and ship a safe correction. Cybersecurity evaluations test model behavior, often through attack-like actions that can create their own exposure if controls fail.

Faster discovery therefore shifts pressure downstream. Triage queues can grow before patch capacity catches up. Evaluation operators need explicit targets, permissions, containment, logging, and stop conditions. Incident responders also need to distinguish a genuine external attack from activity generated by an authorized test.

Chrome’s June result shows what happens when AI reaches the full remediation pipeline. Anthropic’s investigation puts attention on a different threshold: whether evaluation workflows can keep simulated offensive activity inside their approved scope. Security leaders now have to scale both systems without treating them as the same job.

03One Sentence Is Enough to Fabricate a Satellite Scene in Google Earth

A single sentence can now turn Google Earth into a staging surface for synthetic satellite scenes. According to 404 Media, one prompt placed refugees near the Mexican border. The same session added a nuclear plant in Iran and depicted a fatal crash on an Amsterdam street.

The change extends beyond better image generation. It places fabricated scenes inside a map product whose visual conventions already carry authority: overhead perspective, recognizable roads and plausible spatial relationships. Synthetic imagery can now arrive wrapped in an interface that users associate with geographic evidence.

Visible errors have provided a weak backstop, but two research projects show where model builders are concentrating. VideoCoCo targets text-to-video systems that still fail at physical consistency over time. Its authors propose executable code as an intermediate representation to control a scene’s complete spatial and temporal process.

MPIE-Bench targets another common tell: impossible anatomy when several people touch. Its authors list fused limbs, invented extremities and interpenetrating bodies during actions such as carrying or grappling. They also report that vision-language-model judges can saturate interaction scores while human viewers still see geometric failures.

Neither paper validates anything created inside Google Earth. They address two cues viewers often use to reject synthetic media: coherent motion and plausible human contact. As those cues improve, spatial plausibility inside a familiar map frame carries less evidentiary value.

That changes the first step of verification. A believable road layout or natural body movement can demonstrate internal coherence, but neither establishes when or where an image was captured. Media outlets and open-source investigators must trace the origin, inspect available file history and seek independent imagery before treating a map-interface capture as evidence.

The next threshold is no longer whether viewers can spot a malformed limb. It is whether platforms preserve enough provenance for outsiders to separate generated scenes from captured imagery after screenshots and reposts remove the surrounding context.