A specialized non-LLM decision model called Jev cleared the classic Game Boy game Pokémon Red and entered the Hall of Fame on September 23, 2026, relying on Anthropic’s Claude Opus 5 behind the scenes to steer its choices through hundreds of iterations.
How Jev Operates Without a Traditional Chatbot Interface
Developed by TypeSafe AI and led by former OpenAI engineer Diogo Almeida, Jev is a System One model released on September 15, 2026. Unlike large language models that generate conversational text, Jev handles specific decision-making tasks within strict boundaries. It receives a text status and a list of true-or-false questions, returning a structured answer complete with a confidence value.
The system is built for speed and efficiency, boasting response times between 70 and 500 milliseconds. Official claims note it operates 40 to 200 times faster than front-line models like GPT-6 Astra while cutting costs significantly. Yet this speed comes with a trade-off: Jev cannot handle open-ended tasks, leaving those broader challenges to conventional language models.
The Grind Through Kanto and Claude Opus 5 in the Coach’s Corner
Tasked with beating Pokémon Red, developer Andrew Boyd set up a live stream across a browser and terminal where viewers logged in via GitHub to watch. Because Jev cannot read game screens or output freeform text, it depended on Claude Opus 5 to monitor game logs and continuously rewrite option lists and descriptions to keep the agent moving.
The run faced severe obstacles. Harness modification logs show 474 total adjustments logged to correct erratic loops. At various points, Jev walked directly into a closed door at Lorelei’s chamber 53 times in a row and retraced its steps up and down a Rock Tunnel staircase 124 times within ten minutes. After initially falling to the champion’s Alakazam after defeating the Elite Four, the system rebounded to conquer the final battle and enter the Hall of Fame on September 23, 2026.
Balancing Token Costs and Decision Frequencies
Running an autonomous agent through an entire role-playing game requires strict resource management. To keep token consumption down, Claude Opus 5 trimmed the incoming text load sent to Jev by roughly two-thirds, swapping out numbers for plain descriptive text. When Jev got stuck in repetitive movement loops, Opus 5 intervened by slowing down the request frequency from once every second to once every six seconds.
Other developers experimented with alternative setups. Christian Mathiesen ran Jev by reading game memory directly without allowing direct write access, while stmonty trained a smaller world model using over 42,000 screenshots on an RTX 3080 Ti. Meanwhile, Anthropic’s own official Claude Plays Pokémon stream using Opus 4.5 had not cleared the game by January, highlighting how pairing a fast decision model with an overseer model changes the speed and efficiency of complex closed tasks.
Claude Opus 5 Performance Across Software and Science Benchmarks
Beyond acting as a game coach, Claude Opus 5 launched as Anthropic’s new default model on Claude Max and its strongest offering on Claude Pro. It approaches the intelligence of Claude Fable 5 while operating at half the price. On programming benchmarks like Frontier-Bench v0.1, Opus 5 outpaces competing models and doubles the task performance of its predecessor, Opus 4.8, at a reduced cost per task.
The model also shows notable gains in scientific and technical evaluations. Life sciences tests reveal performance bumps across structural biology, organic chemistry, and bioinformatics. Specifically, it scores 10.2 percentage points higher than Opus 4.8 on internal organic chemistry benchmarks involving spectroscopy data and 7.7 percentage points higher on protein sequence variation tasks, pointing to broader utility for structured problem-solving alongside its gaming accomplishments.