TL;DR — Gemini 4 Argon is the new Google default ($2/$10 until 1 December), tried against Google's API. Its thinking tokens are counted, streamed replies once, and 3.7 and 3.5 Flash move to 3.8 and 3.6.
Added
- Gemini 4 Argon (
gemini-4-argon), Google's flagship reasoning model,
generally available since 13 October. It is first in the Google picker and the Google default: 3.1 Pro, still a preview, moves down one place, and a config that names it keeps it. The fresh-install provider is still Z.AI. Through OpenRouter it is google/gemini-4-argon (listed there on 14 October, $2/$10). The owner ran it against Google's API on 14 October with their own key (below); the rest comes from Google's own pages, read the same day. - A paid key only. Google does not offer it on the free tier or on the Google AI Pro plan, so a key without billing cannot call it; the picker says so. - One 1,048,576-token window, shared by what you send, its thinking and its answer. - /thinking offers low, medium and high. Its own default is medium, so Auto runs it there; Max sends high, as there is nothing above it, and no level switches thinking off. (The gate was a gemini-3 prefix, which gemini-4-argon fails, so it would have shipped with /thinking hidden.) - Latency is uneven. The same one-line requests took 3.7 s and 29 s at its default effort, and 59.7 s and 3.5 s at reasoning_effort: "high", in two runs the same day (native: 9.7 s and 4.6 s). Effort did not decide it; allow for tens of seconds, and more for a prompt that needs real thinking. - At least 32,768 reply tokens on every tier. Its thinking counts inside max_tokens, a hard limit, and Google says to raise anything lower for it because it "can spend more of the limit on thinking than earlier models". This reaches the task planner, which sent 2,048. - No temperature, top_p or top_k. Google ignores them on it, so Codeep leaves them out. - $2.00 in / $10.00 out per million tokens (thinking counts as output), through 30 November; $4.00 / $20.00 from 1 December. Cached input is 95% off ($0.10, then $0.20), and there is no cache-write charge or long-prompt tier. Estimates use the rate billed now and do not anticipate the rise on 1 December; revisit the row on that date, not before (docs/MODEL_MAINTENANCE.md).
scripts/verify-gemini-argon.mjs, for the owner to run with their own key
(GOOGLE_API_KEY, at most 12 requests, asks y/N first, --dry-run sends nothing). On the OpenAI-compatible endpoint the CLI uses: a plain turn, a tool call with its follow-up, one at reasoning_effort: "high", and the same plain and tool-call requests STREAMED with usage requested in the stream (how the CLI really sends them). On the native generateContent API the macOS app uses: one at thinkingLevel: high with usageMetadata printed verbatim, and a function call replayed four ways: without its thought signature, with Google's documented placeholder, as returned, and rebuilt the way the app rebuilds it (which proves the signature sits where the app reads it). It prints status, finish reason, usage, timing and where each signature sits for each, and which replay Google accepted. --model <id> points every check at another model (--model gemini-3.8-flash), which is how to find out whether Gemini 3.x behaves like Argon.
Changed
- Gemini 3.7 Flash and 3.5 Flash are gone from the picker. Google deprecated
both on 8 October and routes requests for them to 3.8 Flash and 3.6 Flash, which it names as the replacements; an existing setting moves to that model at launch. Their price and context rows stay, so an old session export still reads. Gemini 3.5 Flash-Lite, a different model, is untouched.
Fixed
- A Google model's thinking tokens are now counted. On Google's
OpenAI-compatible endpoint (what Codeep uses for every Google model) the reply's completion_tokens leaves thinking out and there is no completion_tokens_details: the owner's test of Argon at high effort came back as {"completion_tokens":7,"prompt_tokens":47,"total_tokens":207}, so 153 thinking tokens existed only as the gap in the total — and Google bills them as output. /cost, /stats and the usage ledger counted 7 output tokens where Google prices 160 (its pricing page: the output price includes thinking tokens; no bill was read). When a Google reply's total exceeds prompt plus completion, the difference now goes to the completion count. Nothing is added for any other provider (OpenRouter already folds reasoning into completion_tokens, and elsewhere a total may legitimately differ), and nothing twice if Google ever folds thinking in. This is Google-wide, not Argon-only: it applies to any Google model whose reply carries such a gap. That gap was seen on non-streamed replies. Streamed replies carry total_tokens too (the two streamed requests of the second run), but neither had any thinking in it, so a streamed reply WITH thinking was not seen; the fix reads the same field and its streamed tests use a shape written from the non-streamed one. How far back 3.x under-counted is not known either.
- A streamed reply's usage is counted once. Google's compat stream puts
usage on every chunk (both chunks of both streamed requests of the second run, with the same totals), and chat() recorded every chunk that carried it, so a streamed chat reply — compaction summaries, session titles, /plan, /recall — counted twice on Google. It now keeps the last chunk's usage and records it once. (The agent-turn stream already kept the last; OpenAI sends usage on the final chunk only, so nothing changes there.)
- Plain chat waits up to 5 minutes for Argon, not 60 seconds. The chat
mode of the TUI, /plan, compaction summaries, session titles and /recall all go through one chat() request on apiTimeout, a single timer for the whole reply; replies of 59.7 s and 29 s for a one-line prompt were seen on the same day, the first 0.3 seconds under it. Argon alone gets a 300-second floor there (a larger apiTimeout still wins). 300 seconds is the most a non-streamed reply can wait: Node's own fetch drops a connection that is silent for 5 minutes whatever Codeep sets. Agent turns already had 10 to 15 minutes and are unchanged.
Verified
By the owner, on 14 October, against gemini-4-argon with their own key (scripts/verify-gemini-argon.mjs, two runs of 8 and 12 requests, key not in the output; figures in docs/MODEL_MAINTENANCE.md):
- The OpenAI-compatible endpoint accepts it: a plain turn, a call of a
function tool (read_file), its follow-up with the reply replayed as returned, reasoning_effort: "high" with max_tokens: 4096 (391, finish_reason: stop, 153 and 120 thinking tokens outside completion_tokens in the two runs), and the plain and tool-call requests streamed with stream_options.include_usage. The streamed tool call assembles into valid JSON.
- The native API reports thinking apart: `{"promptTokenCount":47,
"candidatesTokenCount":7,"totalTokenCount":205,"thoughtsTokenCount":151}`. Total is prompt plus candidates plus thoughts.
- A function call replayed without its
thoughtSignatureis an HTTP 400
("Function call is missing a thought_signature in functionCall parts"), on Argon and on gemini-3.8-flash in the same test, and it is a 200 when replayed as returned. The signature sits on the functionCall part (content.parts[0].thoughtSignature; 124 characters on a call with no thinking, 332 to 348 on 3.8 Flash). A call rebuilt the way the macOS app rebuilds it — the call and that part's own signature, nothing else — is answered too, on both models, so the placement the app reads is right.
- Google's documented placeholder (
skip_thought_signature_validator) on
the first call is accepted by both models, so history that carries no signature can be replayed.
- Usage rides on every chunk of Google's compat stream (both chunks of
each streamed request, same totals) — the double counting fixed above.
- A thought signature comes back on a plain answer too, and on streamed
chunks (delta.extra_content.google.thought_signature), not only on function calls.
- Latency is uneven: 2 to 60 s for one-line requests across the two runs,
with no pattern by effort.
- Not a run, but a reading of the code: the CLI is not affected by the
signature requirement, because it keeps history as plain text and never replays a function call (the agent loop in src/utils/agent.ts sends the model's own earlier turns as text and tool results as a user message).
Not verified
Still open after those runs, and each is written down so it is not claimed:
- A streamed reply WITH thinking: whether its final chunk carries the
same gap in total_tokens. The streamed requests had no thinking. The native streamGenerateContent usageMetadata the macOS app reads is not covered by any check, nor is where the signature sits in its chunks (the native checks used generateContent).
- What
usagereports for cached tokens. Every request was far below the
4,096 tokens implicit caching needs, so no cache hit was seen: the 95% rate is Google's published price, not a measurement.
- Gemini 3.x other than 3.8 Flash on the native path (3.6 Flash, 3.5
Flash-Lite, 3.1 Pro): the same family and almost certainly the same, not tested.
- How each client reports Google's refusal for a free-tier key: the exact
text and status were not seen.
- Replays with parallel function calls: every run made one call at a
time. Google puts the signature on the first call of a parallel turn only; the macOS code follows that and it is untested live.
- A signature replayed after a model switch (Argon to another Gemini, or
the reverse).
- The macOS
responseshape{"content": …}is what the rebuilt replay
used and was accepted; other shapes were not compared.
- That 32,768 is the right cap in practice — it is Google's advice, not a
measurement.
Left alone
Google says hard prompts "can take several minutes" and to allow HTTP timeouts of at least 15 minutes. Agent turns here get 10 minutes (12, then 15 later in a run) as one timer that streaming does not reset, plain chat gets apiTimeout (60 seconds; 300 for Argon, above), and Node's fetch gives up on a connection silent for 5 minutes whatever Codeep allows — so 15 minutes would need a new dependency, and a retry after a timeout bills the thinking again. The task planner makes its own request with no timer of its own.