Release Velocity of Coding Agents: Codex Ships 134 Releases in 30 Days

Yesterday I counted the star league of coding agents, today I counted what those projects actually ship. A radar note from September 25 had set this up: the OpenAI Codex CLI sat at release number 1,145, its code had been pushed the same day, and the open question was whether stars and release velocity even tell the same story. The answer, from my own measurement on September 27, 2026 at 07:17 CEST, is no, and the gap is more instructive than any ranking. The Codex CLI now lists 1,158 releases, 134 of them in the last 30 days, more than four per day. The Gemini CLI, meanwhile, lost two stars since yesterday, 107,164 instead of 107,166, and still published 38 releases over the last month. Two curves pulling apart: attention on one side, delivery on the other. If you work with these tools, read both, because stars tell you who is being watched and releases tell you who is shipping.

Method first, because velocity without counting rules is just an opinion. I queried the GitHub API releases endpoints for all seven projects with an authenticated token and counted three things per repository: total published releases, releases in the last 30 days, and releases in the last 90 days, all as of September 27, 2026, 07:17 CEST. Star counts and the last code push came from the repository records, and commits over the last seven days came from the search API: 443 for Codex, 97 for pi. One limit belongs on the table up front: I count whatever GitHub lists as a release, including alpha tags, preview builds and nightly builds. The Gemini CLI necessarily ranks high in the 30-day figure because of its nightly channel. A release is also not a quality verdict: what each release contains lives in the changelog, not in the counter. The measurement is still honest because it is checkable. Anyone can run the same query today and compare numbers.

The table shows the release figures from my measurement, sorted by the 30-day count, with stars next to them for comparison against yesterday's league. The star movement column covers just the one day since September 26 and is deliberately small, because one day is not a curve.

The number needs a translation, because four releases a day sounds like marketing. It is the opposite: a project that writes 134 release titles a month distributes its changes in small portions, and that is precisely what keeps failures small. The lifetime math makes the trend visible. Since the repository was created on April 13, 2025, the average works out to just over two releases per day; the last month ran at more than four, so the cadence has accelerated over the year rather than merely holding. The search API counts 443 commits in seven days, which is sustained operation, not release tourism. The language statistics add 70.4 megabytes of Rust against roughly 1.7 megabytes of Python, the CLI is about 97 percent Rust, exactly what the September 26 radar note suspected.

The latest tag reads rust-v0.159.0-alpha.8 and appeared the morning of the measurement day, and the alpha in the name is part of the truth: part of those 134 is pre-release delivery to the brave, not finished product. That is exactly why I read the number as an operating style, not a quality verdict. A cadence like this means changelogs become your primary reading, every version is outdated within days, and the user of such a harness becomes the version manager whether they like it or not. The MightyBot report, which has named GPT-6 Astra as the bundled default model of the Codex CLI since September, I treat deliberately as a secondary source with an unverified primary trail: it fits the picture of a project shipping weekly, but a report is not a measurement, and that distinction is the point of this entire article.

The most striking number in the table sits with the Gemini CLI: 38 releases in the last 30 days, 110 in 90 days, yesterday's nightly build listed, and at the same time two stars gone since yesterday. Reading stars only, the project looks asleep. Reading releases only, it looks like one of the most active players in the field. Both readings are wrong, and the truth sits between the curves: the project ships in continuous operation yet wins no measurable attention back, at least not in a visible star tick. pi shows the same phenomenon in reverse: only 7 releases in 30 days, but 97 commits in seven days, latest tag v0.87.1 from September 22, and stars up 205 since yesterday. pi develops loudly, publishes bundled, and collects observers; Codex publishes in a steady stream. There is no correct direction here, there are two release cultures, and choosing between them is an operational decision, not a matter of faith.

My deck runs a third release culture, and it is the most important lesson of this measurement for my daily work. The Hermes Agent stands at 36 releases total, 7 in the last 30 days, and its patch release 0.21.4 from September 21 rolls about 1,800 merged pull requests into a single tagged state according to its own changelog. That is the opposite of the Codex cadence: rare, bundled, with curated notes instead of 134 individual changelogs. On an Android phone this is not a matter of taste but of survival math. Every update cycle is a risk, and my anchor is the 88-hour record from my own operation, August 28, 19:40 to September 1, 12:18 without interruption, then zero OOM kills since hardening. A stack built to survive kills only wants releases when they fix a demonstrable defect, which is exactly why I run pi pinned at 0.87.1 instead of daily-current. The 36 of 36 tests in my verified projects were produced by the local pi workflow, Hermes plans, pi executes, Hermes verifies, not through any cloud service. Velocity is an offer, stability is a decision.

The limits, explicitly. My count takes GitHub at its word: every published release counts, including the alpha tag that appeared the morning of the measurement, including the Gemini nightly, and a release without a changelog note counts in this curve like any other. Channel-pure counting, stable tags without pre-releases, is the next measurement step and is still open, which makes these figures a gross measure, not fine-tuned. One cutoff is not a trend: only a second cutoff in October turns two values into a direction, until then every claim about acceleration is tied to a single comparison point, the average over the repository's full lifetime. The commit figures from the search API are a sample, not a full accounting, and star movement over a single day is deliberately marked as a side note. Anyone building a market forecast from this table skips the measurement limits, and I deliberately build none.

October brings the repeat, same seven repositories, same procedure, fixed cutoff, plus channel-pure counting as a second column. Three questions will lead the next run: Does Codex stay above four releases per day, or does the cadence normalize after the alpha surge? Does the Gemini CLI drop out of the top group in channel-pure counting, with the nightly driving its 30-day figure? And does pi keep up the combination of 97 weekly commits and few releases, or does a bundled major land that confirms the star curve? Every number in this article comes from public endpoints, the query takes minutes, and the cutoff is in the first line of this section.

About the author: I am Marcel, a graphic designer and the operator of the HUNTER cyberdeck built on a Google Pixel 6a. I pulled the release and commit figures in this article myself from the GitHub API on September 27, 2026 at 07:17 CEST, the star baselines come from my measurement of September 26, 2026. The operating anchors live in the vault: stability-snapshots for the 88-hour series, the radar notes since September 25. Last reviewed on September 27, 2026.

Sources: GitHub API queries against the releases endpoints and the language statistics of openai/codex, earendil-works/pi, google-gemini/gemini-cli, sst/opencode, NousResearch/hermes-agent, openclaw/openclaw and herdrdev/herdr, all fetched September 27, 2026, 07:17 CEST, authenticated. GitHub search API for commits over the last seven days, same fetch. Hermes Agent v0.21.4 release note of Sep 21, 2026 with about 1,800 merged pull requests, GitHub releases page. My own star measurement of September 26, 2026, 07:03 CEST, radar in the vault. Radar notes of Sep 25 and 26, 2026 on Codex v0.157.0 as release number 1,145. The MightyBot report on coding agents, deliberately marked as a secondary source with an unverified primary trail. Marc Bara, Hermes Agent build log, Medium, read September 27, 2026.