Haiku 5.5 takes the WebMCP crown from Jev
Idan Levin's benchmark now puts Haiku 5.5 with WebMCP ahead of previous winner Jev on combined success, speed and cost, while OpenRouter shipped a production tool call guard running on Jev 1.13. Two results, opposite directions, same week.
Built with it
vogel says Jev 1.14's new bounding boxes work for computer use
The first thing anyone did with the bounding box type from Jev 1.14 was point it at a screen, and vogel says it is pretty accurate at computer use. The release from Diogo Almeida also turned on image input and context up to 262k, both behind opt-in TypeSafe-Danger-Zone headers, with OpenRouter listing the price unchanged at $0.042 per million input tokens and output free.
Yep, the Jev you know and love was an image model this whole time.
Someone measured it
Haiku 5.5 with WebMCP took the top spot from Jev
Idan Levin added six models to the WebMCP benchmark and says Haiku 5.5 using WebMCP now takes first on the combined success, speed and cost score, ahead of GPT-6 Luna and previous winner Jev. Against the top computer use configuration, Luna, Haiku solved 49 of 49 tasks versus 41, at 3.7 seconds per task versus 19.4 and 8.4 times lower cost.
haiku 5.5 is ~3.5 seconds per task, which is more than 2x faster than other models with webmcp
SubQ Snap's 61.7 against Jev's 57.9 is still only Whedon's number
Alexander Whedon launched SubQ Snap claiming 61.7 on Decision Index v0.2.1 against Jev's 57.9, with 384K multimodal context, 660ms at 100k tokens and $0.05 per million input tokens, output free. Nobody has published an independent run since. Moritz Kremb had early access and built a Chrome extension that describes his screen and clicks for him, but posted no numbers of his own.
Snap is the only decision model or LLM where compute per token stays constant as context grows.
OpenRouter says Jev 1.13 beat Sonnet 5 on its Armor test set
OpenRouter's Armor beta scores each tool call against a rule set plus a slice of the transcript before an agent executes it, and it runs on Jev 1.13. On an internal test set OpenRouter generated for the task, it says Jev 1.13 had the highest accuracy and lowest cost against other Decisions models and Claude Code's usage of Sonnet 5. @completeskeptic reposted it with "SO FKIN COOL".
Armor makes Claude Code's auto permissions mode compatible with any model.
Writing criteria
Lee Knowlton's rule: build the verifier, or work one task at a time
Lee Knowlton sorts work into AI-verifiable and not: if a model can independently tell whether its own output is good, build the verifier and stop being the bottleneck, otherwise stay serial and protect your focus. The middle case is an AI judge, which he says only counts once you have evals proving the judge is competent, and he points at Mike Taylor for that part.
The most valuable thing you can do right now is find a way to make an unverifiable domain verifiable by AI and let it rip.
Also on the timeline
Jev images land in MotherDuck SQL
Hamilton Ulmer says Jev image support in MotherDuck SQL matches Sonnet quality at ten times the speed and a 25th of the cost. Confidence scores rescued a weaker benchmark.
Open SWE router: $0.94 a thread against $2.61
LangChain's Open SWE coding agent now routes each thread to the cheapest model likely to finish the job, and @neil_xbt's summary of the test put median cost at $0.94 a thread against $2.61 for always using GPT-6 Astra, a 64% drop.
Frequently asked questions
Does Jev 1.14 work well for computer use?
Yes, according to vogel, Jev 1.14's new bounding boxes (outline boxes around objects on screen) are pretty accurate at computer use, though no specific accuracy figure was published.
What new features came with Jev 1.14?
Jev 1.14 added bounding box type, turned on image input, and increased context up to 262k tokens (the amount of previous conversation the model can remember), with both images and long context available behind opt-in TypeSafe-Danger-Zone headers as preview features.
Which model won the WebMCP benchmark?
Haiku 5.5 using WebMCP (a browser automation tool) took the top spot on the combined success, speed and cost score, ahead of GPT-6 Luna and Jev.
How much does Jev cost on OpenRouter?
OpenRouter lists Jev's price at $0.042 per million input tokens with free output.