One ternary in Jev's gateway limits it to hinting inside Claude Code

Anthropic's API forces the hint-only design. According to the Anthropic docs the post cites, forcing a tool with extended thinking on returns an error, and on the newest models a forced tool returns a 400 regardless of thinking ^[8]. Changing tool_choice invalidates the messages cache. Editing a tool definition invalidates everything ^[9].

The gateway touches neither. In hint mode it appends a block after everything the client sent, on the client's own blocks, so the cached prefix stays byte identical ^[7]. The block is a `<system-reminder>` saying a tool-routing model suggests a tool, worded so the model knows it may disagree ^[7]. Given those constraints, I think this is the correct design. A guard skips the hint entirely when the last block is already a system reminder, and Claude Code appends those itself ^[10]. Some turns therefore reach the model with no suggestion at all ^[10].

Jev is TypeSafe's decision model. It returns a choice, a score or a yes-no probability in 70 to 500 ms, at $0.042 per million input tokens ^[2]. The install is clean: `bin/clients.mjs` sets `ANTHROPIC_BASE_URL` and nothing else, so a Pro or Max login keeps working without an API key ^[5]. The gateway's README says: "Expect better tool picks on large tool lists, not lower cost or latency." ^[6] The post counted six videos in five days that called the setup the cheapest agentic coding loop available ^[1].

The bench, jev-gateway-bench, is 120 sessions at five runs per cell on Claude Code 2.1 ^[11]. On chess-san, a feature task, Opus 5 with routing on also made 47% more requests and took 83% longer ^[13]. Sonnet 5 read 16% more input tokens and took 37% longer ^[14]. The authors' explanation, as the post relays it, is that a hint that does not fit costs a detour instead of being ignored for free ^[15]. One model solved the feature task five times out of five alone and three out of five with routing on, using half the input tokens ^[16].

The post's headline says Jev did not make Claude Code cheaper ^[24]. That holds on the feature task. The bug-fix gains came from the same bench, run inside Claude Code where the gateway can only hint, so hint mode did cut tokens on that task ^[1].

The router's own input is a separate cost. The gateway re-sends the tool roster on every call, and an open issue covers that behaviour ^[19]. The author's working setup sent 40 tools and 57,000 prefix tokens ^[18]. At Jev's rate that is about $0.0024 per turn ^[2], with roughly 14,000 tokens read for each token answered ^[3]. A bench footnote records one setup sending 285 tools and about 200,000 tokens with every Claude Code request, against 6 tools and 7,000 tokens clean ^[20]. The author calls Jev's pricing survivable at those sizes ^[21].

The evidence is narrow by the author's own account: one benchmark, two task families, five runs per cell, on a chess engine, with proxy counts from one machine and one set of plugins ^[22]. For the bug-fix column to transfer, a user needs a small roster and work that looks like bug hunts, and the author points to the bench repo for measuring that locally ^[23].