Jev took a median .35 seconds per passage in one outside test. The comparison model took 8.83 seconds.
That made Jev roughly 25 times faster in that test. But Jev caught six of seven intended defects, while Fable 5.1 at high effort caught all seven.
Speed is useful. A valid answer is useful. Neither guarantees the right decision.
Jev Was Faster, but Missed a Defect
Typesafe introduced Jev in September 2026, claiming up to 200 times faster performance on decision tasks. That speed comes from narrowing what the model can do.
Jev cannot write an answer. The application supplies the state it may inspect and a set of allowed answers, then Jev chooses among those options and returns probabilities. If important evidence is missing from the supplied state, Jev cannot inspect it.
That structure may fit bounded decisions such as routing a support ticket, selecting a tool, or rating urgency. It also keeps the output inside the required schema.
Typesafe calls this “zero hallucinations.” The guarantee applies to the response format, not factual correctness.
Typesafe reports calls taking roughly 70 to 500 milliseconds. Across four workflows, it reported 67.8% aggregate agreement, including 76% for customer service. These are vendor results, and the reference answers came from two larger models rather than a human review against the relevant business policy.
The outside test shows the engineering tradeoff more clearly. Jev was much faster, but it missed one intended defect that Fable 5.1 at high effort caught.
The practical evaluation is narrow: run a repeated decision in shadow mode, keep ambiguous cases for review, and leave execution behind explicit, code-enforced boundaries. Jev may choose among allowed tools, but ordinary code should still decide whether a user is authorized or whether a payment exceeds a fixed limit.
First Audio Is Not a Complete Latency Measure
A carefully configured local voice agent ran on an RTX 3060 with 12 gigabytes of graphics memory. In the reported warm test, it generated 4.88 seconds of speech in 3.05 seconds, with first audio arriving at 0.94 seconds.
That measures warm speech generation, not the complete delay between a question and a correct answer. A voice agent also waits for turn detection, transcription, tool calls, buffering, playback, and possibly a cold start.
A teaching example placed a fully sequential pipeline at 6.8 seconds before playback. Reusing an early transcript and starting speech synthesis after the first complete answer sentence reduced that example to 3 seconds.
Those timings were illustrative, not GPU measurements. The supported lesson is about scheduling: improving the pipeline can matter as much as measuring one model.
Memory is shared too. The speech process reportedly used about 4.3 binary gigabytes of graphics memory, while the language model, caches, buffers, operating system, tools, and offloaded weights also needed room. Hugging Face’s unquantized Nvidia example budgets 24 GB of graphics memory plus system RAM, so a 12 GB setup requires deliberate choices about model size, quantization, and CPU work.
Pythagoras clears stale speech queues and rejects late output from an older run. That can stop an obsolete answer from playing, but it cannot reverse a tool action that already happened.
Spending Agents Move Controls Earlier
Stripe says Link, its wallet with about 300 million consumers, has quickly become a wallet for agents. That creates control problems for both the person authorizing the spending and the business accepting the purchase.
Emily at Stripe described spending thresholds and approvals as the initial control. An agent can operate within constrained parameters, then request approval when cost or customer impact crosses a boundary.
Fraud can begin before payment. Cursor faced attackers creating accounts, taking free credits, abusing trials, and accumulating unpaid usage. Because inference has a real marginal cost, Stripe moved fraud checks earlier, including account creation, trial activation, and usage accumulation.
Pricing creates another unresolved boundary. Token billing can follow the underlying model cost, but outcome pricing becomes harder when customers assign different values to the same product.
Stripe also describes evaluations helping a router choose among quality, cost, latency, and reliability. That approach is only as useful as the evaluations behind it.
Coding Agents Need More Than Orchestration
More orchestration features do not automatically produce more completed tasks.
A Berkeley study compared 7 models across Claude Code, Codex, and Pi. Its reported conclusion was that the harness barely changed success rate, but did change cost. Across 21 model and harness pairs, the model’s native command-line tool was not always the best option.
In that evaluation, workflow and cost were practical reasons to choose a harness. Matching the harness brand to the model did not guarantee the best result.
The operational work extends beyond the model: coordination, planning, reviews, worktrees, history export, quotas, capacity, bugs, secrets, and security flaws. One open-source tool warns that exported local assistant histories may contain secrets. A security write-up also reported a Plugin4Shell zero-click remote-code-execution flaw through marketplace auto-updates, with Claude Code and Codex described as patched.
The useful target is not the fastest response. It is a correct, tool-backed result inside explicit approval, execution, and reversibility boundaries.