The usable context story in Antigravity just changed on paper, and the paper needs reading with care. Antigravity Pulse, a monitoring tool built for Google Antigravity, released v1.3.2 with new starting baselines of 512,000 tokens for Gemini 4 Argon, Gemini 4 Carbon, Gemini 4 Hydrogen and Orchestrator. That figure is double the 256,000 token point that dominated discussion of Antigravity all month.
Google has not announced that it removed the old clamp. That distinction is the whole story. Pulse now reads live ceilings from the session itself and starts its tracking at 512K for the Gemini 4 models. Whether every Antigravity session now gets 512K before compaction is a claim no Google page checked for this report confirms.
The Announcement
There was no Google announcement with this headline. The release that carries the numbers is Antigravity Pulse v1.3.2, covered by Undercode News on November 16 from the Pulse changelog. Pulse is a separate monitoring project for Antigravity users. It is not a Google product, and its changelog is not a Google promise.
What Pulse says it did is still notable. The tool now inspects native GeneratorMetadata, along with other session signals, to detect the actual ceiling a backend enforces for a given session. The changelog lists ceilings it can adapt to at 80,000 tokens, 256,000 tokens, 512,000 tokens, 840,000 tokens, and 1 million tokens and beyond. It also filters out raw catalog entries of 1,048,576 tokens, because a catalog entry does not prove what a live session received.
Context is the amount of text, code, tool output and history a model can hold in view at once. A token is a chunk of text the model counts as it works. Compaction is the step where the harness summarises older turns and drops bulky raw history so a session can continue. Compaction keeps a session alive. It also loses detail, and that loss is why the trigger point matters so much to people running long coding sessions.
Important Details
The numbers come from three places, and they do not carry the same weight. The table keeps them apart.
| Item | Figure | Source |
|---|---|---|
| Advertised Gemini 4 window in Antigravity | 1,048,576 tokens | Client registration and /context view, as documented in the November 5 forum report |
| Default compaction point, App 2.29.0 and CLI 1.5.1 | About 250,000 to 256,000 tokens | Google staff reply in the same forum thread |
| Model calls in the forum evidence archive | 15,575 calls across 209 sessions, all at 256,000 max context | Forum report, November 5 |
| Pulse v1.3.2 baseline, Gemini 4 models and Orchestrator | 512,000 tokens | Pulse changelog, via Undercode News |
| Pulse v1.3.2 baseline, Gemini 3.8 Flash and Claude 5.5 | 256,000 tokens | Pulse changelog, via Undercode News |
| Pulse v1.3.2 baseline, GPT-OSS 120B | 80,000 tokens | Pulse changelog, via Undercode News |
From the old clamp to the new baseline, tokens
The 256K bar is the default compaction point Google staff described for CLI 1.5.1. The 512K bar is the Pulse v1.3.2 starting baseline for Gemini 4 models. The 1,048,576 bar is the advertised window. These are three different kinds of figure, shown together only for scale.
The forum evidence behind the old figure is unusually detailed. The November 5 report on the Google AI Developers Forum says the shipped binaries capped the compaction threshold at 256,000 tokens in code, at about 25 percent of the advertised window, for every model. It cites session data to match, and it notes that each compaction stalled a session for a median of 70 seconds and discarded a cached prefix of more than 240K tokens.
Google staff replied in that thread with a narrower claim. They said automatic compaction in App 2.29.0 and CLI 1.5.1 triggers by default around 250K to 256K tokens to maintain latency, reliability and token efficiency, and is not currently user configurable. A day earlier, in a separate feature request thread, staff used similar language and passed a request for a user selectable size to the engineering team.
Read those two replies together and the old position is clear. Google described 256K as a default. Users who read the binary said a default you cannot raise is a limit. Both descriptions point at the same practical fact for the versions named. Sessions compacted near 256K.
What Pulse Changed About Compaction
The second half of the headline is about reporting, not about Google code. Pulse v1.3.2 unifies manual and automatic compaction under one wording, Context Compacted. It also expires stale compaction events automatically.
That sounds minor until you picture the failure it fixes. A stale notice could sit in the tool and hide a new warning when a session crossed 85 percent of its ceiling. With stale events expiring, a fresh warning can show when it should. Pulse also adds strict separation between quota pools on a five hour cycle, so usage from one model pool is less likely to be mixed with another.
Google has been working on the same pain from inside the product, on a slower public record. The official Antigravity changelog for CLI v1.4.10, dated October 29, says the after compaction size shown in the marker and in /context now counts what the summary carries into the next prompt, so the number is closer to the real size. An earlier CLI release, v1.2.8 on September 22, says compaction budgets are now sized from the full context window of the model instead of the compaction trigger threshold, so summaries are no longer cut short early. Those are real improvements to how compaction behaves and reports. Neither entry says the 256K trigger was raised or removed.
We covered the wider Antigravity workflow picture in Run Antigravity Workflows With Parallel Agents Safely, and the newest fast model for those workflows in Gemini 4 Hydrogen Fast Agent Workhorse Tool. Both pieces rest on the same practical question this one does. How much working memory does an agent get before the harness starts summarising its past?
What’s Next
Confirmed steps are short. Pulse v1.3.2 is out, with the baselines and the detection method described above, according to its changelog as reported by Undercode News. The official Antigravity changelog checked on November 16 runs to CLI v1.5.8 on November 13, which adds Gemini 4 Hydrogen and Carbon support and context usage updates for editor sessions. That entry does not announce a new compaction threshold.
Open questions are the ones that decide whether the headline becomes a Google fact. Will Google publish a new default threshold in its own changelog or docs? Will /context show the real trigger for a session, as users asked on November 5? Will the 512K and 900K options that the forum report found gated in the binary, at about 1.3 times and 1.8 times quota per turn for Argon, open to subscribers, and at what price in quota?
- September 22, 2026Antigravity CLI v1.2.8 sizes compaction budgets from the full context window of the model, per the official changelog
- October 29, 2026Antigravity CLI v1.4.10 improves after compaction size reporting in the marker and in /context, per the official changelog
- November 4, 2026Google staff call the about 250K compaction threshold a fixed default that is not user configurable yet, in a feature request thread
- November 5, 2026A forum report documents a 256,000 token clamp in Desktop 2.29.0 and CLI 1.5.1, with session data from 209 sessions
- November 13, 2026Antigravity CLI v1.5.8 adds Gemini 4 Hydrogen and Carbon and context usage updates for editor sessions, per the official changelog
- November 16, 2026Undercode News reports Pulse v1.3.2, with 512K baselines for Gemini 4 models and an overhaul of compaction event reporting
Reporting has improved in public, step by step, since September. The trigger is where the public record still stops at a staff reply from early November.
So treat the doubling as reported, not as confirmed by Google. Pulse is a useful witness because it reads session metadata instead of trusting a catalog. It is still a witness, not the product team. If your own Antigravity sessions now run past 256K before the first summary, that is the evidence that matters, and Pulse v1.3.2 is built to show it. If they still stop near 256K, the new baseline is a better label on the same wall, and Google still owes its subscribers a straight answer in its own docs.
Facts in this report were checked on November 16, 2026 against the Google AI Developers Forum threads cited below, the official Antigravity changelog, and the Pulse v1.3.2 changelog as reported by Undercode News.
Sources
- Google AI Developers Forum, 1M advertised context vs. a 256K compaction clamp in the shipped binaries, November 5, 2026. Source for the binary clamp, the session data, the 512K and 900K gated options, and the Google staff reply on the 250K to 256K default.
- Google AI Developers Forum, Enable larger than 250K Gemini context sizes in Antigravity Desktop, November 4, 2026. Source for the staff statement that the threshold is a fixed default and is not user configurable yet.
- Undercode News, Antigravity Pulse v1.3.2 Just Made AI Context Tracking Smarter, November 16, 2026. Source for the Pulse baselines at 512K, 256K and 80K, the live ceiling detection, the quota separation, and the compaction wording and stale event changes.
- Google Antigravity, Changelog, checked November 16, 2026. Source for CLI v1.5.8 on November 13, CLI v1.4.10 on October 29 on after compaction size, and CLI v1.2.8 on September 22 on compaction budgets sized from the full window.
- Hero image: Dirty Static, generated in house style. Body image: Dirty Static, generated in house style.
END OF TRANSMISSION
Stay curious. Question everything.


