MAY 22 2026 / 42 COMMITS / 2 RELEASES

New AI on May 22, 2026. No frontier lab shipped. The dated record worth opening was one open-source agent making Gemini actually usable, in a four-argument function signature.

Neither Hugging Face nor GitHub publishes a list keyed to a calendar day, and on May 22, 2026, none of OpenAI, Anthropic, Google DeepMind, Mistral, or Meta cut a dated frontier-model drop. So a query like this one rarely produces what it literally asks for. What it can produce is one open-source project whose commit log carries the exact date 2026-05-22, and one decision inside that day that explains something about where desktop agents are.

Matthew Diakonov, Written with AI

Published May 24, 20269 min read

Direct answer, verified 2026-05-24

There is no canonical "new AI on May 22, 2026" list because none of the major model labs cut a release that day, and the discovery platforms do not index by date. The honest answer is three live feeds:

huggingface.co/papers for research where an implementation is linked.
huggingface.co/models sorted by trending, for weights and quantized variants.
github.com/trending for the application layer of agents, MCP servers, and inference engines.

For one verifiable open-source AI release dated 2026-05-22: the macOS agent Fazm shipped v2.9.35 and v2.9.36. The keystone change was a one-line signature update in acp-bridge/src/gemini-query.ts that made Google Gemini Flash and Gemini Pro able to call the agent's MCP tools (screen capture, browser, WhatsApp, the Assrt browser-agent MCP). On the same day, Gemini also became a free fallback option in the app when built-in Claude credits ran out. Every fact below is traceable to CHANGELOG.json and git log on the repository.

Why a calendar-day query and a trending feed disagree

A trending score is built to surface popular work, not new work, and it has no notion of a date. The same model can hold the Hugging Face top slot for a week. A paper that arrived three days ago can outrank a release that landed this morning. If your question is literally "what shipped on this exact day," a homepage feed is the wrong instrument. The right instruments are dated project changelogs and per-day commit logs. They are less social, but the timestamps are real.

That is why the rest of this page narrows in on one project. A list of ten trending repositories would tell you what was loud that week. One project's May 22, 2026 tells you what shipping actually looked like that day, including the small decisions that aggregator feeds will never index but that determine whether a published model is actually usable inside an agent loop.

One open-source repository's actual May 22

Fazm is an open-source macOS app that wraps Claude Code (through the Agent Client Protocol) and Codex in a native UI, with persistent sessions, one-click forking, and no auto-compacting of context. The repository lives at github.com/mediar-ai/fazm. On 2026-05-22, 42 commits hit main and two production releases were cut. The density of that day is the texture a trending feed cannot show:

commits to main between 00:00 and 23:59 Pacific

production releases (v2.9.35 at 13:43 UTC, v2.9.36 at 19:27 UTC)

argument added to buildMcpServers that closed the Gemini gap

Two threads ran through the day. The morning thread, around 11:09 to 11:35 Pacific, was the credit-and-pricing plumbing that lets the app suggest Google Gemini as a free fallback when bundled Claude credits run out. The afternoon thread, around 15:10 to 15:15 Pacific, was the keystone batch: a four-commit run that added a single modelId parameter to the function that spawns MCP servers, so that sibling MCP processes could pin their credential providers per session per model.

The git log itself

Run this against a local clone of the public Fazm repo and you get the same answer this page is built on. Eight of the 42 commits land first:

bash

The full set is reproducible with the same command, dropped of the head -8 filter. Author-filter to m13v if you want only human commits; two of the 42 are the Codemagic CI release tags.

The one signature change that closed the Gemini gap

The CHANGELOG entry for v2.9.35 includes one line that does most of the work: "Fixed Gemini Flash and Gemini Pro not being able to call MCP tools (capture_screenshot, browser, WhatsApp, etc.)". The commits behind it are not glamorous. The keystone is commit 747f8893 at 15:12:17 Pacific, which changes a single call site in acp-bridge/src/gemini-query.ts line 95:

acp-bridge/src/gemini-query.ts (line 95)

// before
const mcpServers = buildMcpServers(mode, cwd, sessionKey);

// after (commit 747f8893, 2026-05-22 15:12:17 -0700)
// Pass the model id so sibling MCP servers (Assrt) pin their
// credential provider to Gemini for this session instead of
// inheriting the bridge-spawn default (which is usually Claude).
const mcpServers = buildMcpServers(mode, cwd, sessionKey, modelId);

The matching dispatch lives in acp-bridge/src/index.ts. Commit 64e26ee3, two minutes earlier, refactored the MCP-spawn block so that the Assrt MCP's credential provider is selected based on the modelId the session is using. 43 line diff. Before that block existed, the Assrt MCP got the bridge's default credential provider, which was Claude-shaped. When a Gemini session asked the Assrt MCP to drive a browser, the credential layer rejected the call.

Two other commits in the same 15:10 to 15:15 batch threaded the same argument the rest of the way. 7129614e did the same passthrough for handleCodexQuery (Codex sessions had the same bug, just less visible because Codex did not yet exercise as many MCP tools). e939ce0c added an activeModel field to GeminiQueryDeps so the bridge could keep the choice in scope between buildMcpServers and the streaming loop.

What the Gemini backend could and could not do, day-of

On the morning of 2026-05-22, Gemini Flash and Gemini Pro were already selectable in the model picker. They could stream text. They could not usefully drive anything else. On the evening of 2026-05-22, after the two releases, they had the same MCP-tool surface as Claude. The full matrix:

Feature	Before (May 22 morning)	After v2.9.35 + v2.9.36 (May 22 evening)
Selectable in the model picker	Yes (already selectable before May 22)	Yes
Call capture_screenshot MCP tool	No, tool call failed at credential resolution	Yes, modelId forwarded into MCP spawn
Call browser MCP (browser-harness)	No, same credential failure	Yes
Call WhatsApp MCP	No	Yes
Call Assrt browser-agent MCP	No, Assrt provider defaulted to Claude credentials	Yes, Assrt provider pinned per selected model
Free credit fallback in the app	Not offered, error told you to connect Claude or ChatGPT	Suggested in credit-exhausted message, excluded from cost cap
Personal account chooser entry	Two options (Claude, ChatGPT)	Three options (Claude, ChatGPT, Gemini)

Source: CHANGELOG.json and git log for 2026-05-22 at github.com/mediar-ai/fazm. Verified against a local clone on 2026-05-24.

The day, hour by hour

Reading the commit log in order makes the day's logic visible. The morning lays the credit-and-pricing plumbing that makes Gemini-as-fallback economically possible. The afternoon adds the integration plumbing that makes Gemini-as-fallback technically useful. The evening generalizes the Assrt MCP into a browser agent the parent can offload to, and lands a contextual ad slot with PII redaction:

2026-05-22 commit log, ordered by Pacific time

11:09 PT cce0f754, 570c0e63: /api/stripe/variant-price endpoint goes live. Paywall now reads pricing from a single source.
11:11 to 11:14 PT d58e66a1, 3683956d, 9e4d5441, a92e1855: PaywallSheet gets dynamic pricing and 7-day free trial support, header copy becomes data-driven.
11:17 to 11:20 PT dcc11336, 8ecb7550, 13226b62, 05fd3da2, 2db8f0fb, f57d5854: Personal account chooser learns about Gemini, credit-exhausted error message suggests Gemini, Gemini models are excluded from the builtin cost cap.
11:35 PT 6e464533: CHANGELOG updated for v2.9.36 'Google Gemini as a free option when built-in AI credits run out'.
12:56 PT a836a05e: Sentry noise filter added for the repetitive DB and auth errors generated by the morning's deploy churn.
13:43 UTC (06:43 PT) 73b05a5a: v2.9.35 release tagged on Codemagic CI. CHANGELOG entry: 'Fixed Gemini Flash and Gemini Pro not being able to call MCP tools'.
15:10 to 15:15 PT eae8a87d, 64e26ee3, 7129614e, 747f8893, e939ce0c, e1b12bfb, 1789ae78: the keystone batch. buildMcpServers gains a modelId parameter, GeminiQueryDeps and CodexQueryDeps gain activeModel, the Assrt MCP provider is configured per selected model.
15:22 PT 6a11059f: changelog amended to note the Assrt QA test fix that the credential plumbing unblocked.
15:59 to 16:05 PT 8a72ba2b, 8bcd9027: selected Gemini model id is forwarded into the Assrt environment, Assrt's Gemini default becomes Flash.
17:01 PT d6fd564d, fc912791: chat prompt schema picks up column annotations and a stricter MEMORY.md format rule.
17:48 to 18:06 PT 20a6096d, 0ff02ac6, f7f9a5f9, fd89b03b, 1763c165, 0e3102f0, c0d68b42, 375a95e8, e9036a5b: Assrt MCP is refactored from a QA tool into a general-purpose browser agent, and a 'No browser MCP (Assrt only)' mode is added to Settings.
17:50 to 18:36 PT a0a6c4fe, d712bc88, 498e5f24, 84458627, 4216ef24, bd47e739: Koah ad surface lands (KoahAdView with PII redaction, developer override, click-through routed to the system browser).
19:27 UTC (12:27 PT) 8b83fe83: v2.9.36 release tagged on Codemagic CI. CHANGELOG entry: 'Added Google Gemini as a free option when built-in AI credits run out, alongside Connect Claude and Connect ChatGPT'.

1 line

“The CHANGELOG entry that absorbs the day's load-bearing fix is six words long: Fixed Gemini Flash and Gemini Pro not being able to call MCP tools. The one-line code change behind it is a four-argument function signature in gemini-query.ts.”

Fazm CHANGELOG.json v2.9.35, dated 2026-05-22

What this means for tracking new AI on a specific day

Most days, the answer to "what new AI shipped on this calendar day" is not a model. It is plumbing. Plumbing is not what trending feeds index, but it is the thing that determines whether a model that was already released is usable inside an agent that already exists. Gemini Flash and Gemini Pro were already shipped on May 22. The interesting May 22 event was the day the agent layer made them work with the rest of the tool surface a real agent needs.

If you are trying to track this layer, the cheapest reliable signal is a small set of project changelogs read on a rolling 7-day window. CHANGELOG.json files. Releases pages. git log on main. The aggregator feeds will pick up the announcement post days later, if at all. The dated record was there the whole time.

Where to verify any of this

Open github.com/mediar-ai/fazm and read three dated artifacts. CHANGELOG.json at the root lists every release with its date; v2.9.35 and v2.9.36 are present, both dated 2026-05-22. The matching notarized macOS builds are served from fazm.ai/download. The commits backing every claim above are reachable from main; the shas in the timeline section are stable.

The point is that nothing on this page asks you to take a date on faith. A dated, line-level record is what makes a calendar-day claim something other than a press summary.

Want to walk through how the agent picks a backend?

Book 20 minutes. I'll share the screen and step through the modelId dispatch live, on a real session.

Frequently asked questions about new AI on May 22, 2026

What new AI models, papers, or open-source projects were released on May 22, 2026?

Neither OpenAI, Anthropic, Google DeepMind, Mistral, nor Meta cut a frontier-model release on May 22, 2026. Hugging Face and GitHub do not publish indexes keyed to a calendar date, so a literal 'May 22 list' does not exist there either. The honest, verifiable open-source artifact for that date is the macOS agent Fazm: 42 commits to main, two production releases (v2.9.35 at 13:43 UTC and v2.9.36 at 19:27 UTC, both stamped 2026-05-22 in CHANGELOG.json), and a single change that did the load-bearing work, which was making Google Gemini Flash and Gemini Pro able to call MCP tools. Every fact on this page is traceable to git log and CHANGELOG.json at github.com/mediar-ai/fazm.

Where do I actually look for new AI in a 24-hour window like May 22, 2026?

Three rolling feeds, all ranked by a trending score, none calendar-keyed. huggingface.co/papers/trending for research with linked implementations. huggingface.co/models?sort=trending for weights and quantized variants. github.com/trending for the application layer of agents, MCP servers, and inference engines. To verify whether a specific project actually shipped on a specific day, you have to open its own dated changelog. The Fazm releases for May 22, 2026 are in CHANGELOG.json at the root of github.com/mediar-ai/fazm, with version, date, and changes fields per release.

What did Fazm actually change about Gemini on May 22, 2026?

Two things, on the same day, in the right order. Earlier in the day (around 11:09 to 11:35 Pacific), the credit-exhausted error message learned to suggest Google Gemini as a free alternative when the bundled Claude credit pool ran out. The PersonalAccountChooserSheet got a Gemini option, and Gemini models were excluded from the builtin cost cap. Then in the afternoon (15:10 to 15:15 Pacific), a sequence of commits added a modelId parameter to buildMcpServers() in acp-bridge/src/index.ts and threaded it through handleGeminiQuery and handleCodexQuery. That was the keystone fix. Before it, the Gemini backend could not call any MCP tool (no screen capture, no browser, no WhatsApp). After it, it could. The CHANGELOG entry that summarizes this is one line: 'Fixed Gemini Flash and Gemini Pro not being able to call MCP tools'.

Why was that one signature change load-bearing?

Because sibling MCP servers run as separate processes that the bridge spawns, and they each need a credential provider configured at spawn time. The Assrt browser-agent MCP, for example, has to know which model the parent session is using so it can pin Assrt's own credential provider to that model. Before May 22, buildMcpServers(mode, cwd, sessionKey) did not carry the active modelId, so spawned MCP servers inherited the bridge default (usually Claude). When a Gemini session tried to drive an MCP tool, the tool call routed through credentials that the Gemini backend could not satisfy and the call failed. After buildMcpServers(mode, cwd, sessionKey, modelId), each session's MCP servers pinned their providers per-model. Five lines added across acp-bridge/src/gemini-query.ts and acp-bridge/src/codex-query.ts, plus 43 lines in acp-bridge/src/index.ts where the dispatch lives.

Where is the canonical record of these commits?

Two on-disk sources, both public. The first is CHANGELOG.json at the root of github.com/mediar-ai/fazm. The v2.9.35 entry is dated 2026-05-22 and lists the MCP-tool fix. The v2.9.36 entry is dated 2026-05-22 and lists the credit-pool addition. The second source is git log itself. The four most load-bearing commit shas on that day are 747f8893 (Pass modelId to buildMcpServers to pin Gemini credentials), 64e26ee3 (Configure Assrt MCP provider based on selected model), e939ce0c (Pass active model to buildMcpServers for Assrt credential pinning), and 7129614e (Pass modelId to buildMcpServers in handleCodexQuery). All four landed between 15:10:50 and 15:15:15 Pacific. The notarized macOS builds for v2.9.35 and v2.9.36 are served from https://fazm.ai/download.

Did the major AI labs announce anything notable on May 22, 2026?

No dated frontier-model drop from OpenAI, Anthropic, Google DeepMind, Mistral, Meta, or the major Chinese labs landed on May 22, 2026. The May 22 activity worth recording is at the agent and tooling layer, where the gap between a published model and a model that can actually drive a Mac is still measured in days. Gemini Flash and Gemini Pro had been available via API for some time. May 22 is the day they became a one-click backend inside a native Mac agent that can also call the agent's MCP tool surface (screen capture, browser, WhatsApp, accessibility-driven Mac control, the Assrt browser-agent MCP).

What else shipped at Fazm on May 22, 2026 besides the Gemini work?

Two other changes worth noting, both in the evening Pacific. First, the Assrt MCP was refactored from a QA-testing tool into a general-purpose browser agent (commits e9036a5b, 375a95e8, c0d68b42 at 18:04 to 18:06 Pacific), which means the parent agent can hand off a multi-step browser flow to a sibling Assrt agent with its own browser session instead of driving it step by step. Second, a 'No browser MCP (Assrt only)' mode was added to Settings (commits 1763c165, 0e3102f0 at 17:53 to 17:54), which collapses the browser-tool surface to just Assrt for sessions that do not need direct browser control. None of this required a model release. All of it required real plumbing.

Why does this read as 'one open-source repo' rather than a broader roundup?

Because a broader roundup for a specific calendar day is dishonest when no major lab actually shipped on that day. Pretending otherwise produces the same generic 'AI news of the week' page everyone else writes. A single project's commit log on the day is a smaller scope and a higher confidence answer. It also surfaces the kind of integration-layer work that aggregator feeds never index but that determines whether a published model is actually useful inside a real agent loop on a real machine.

What is the right cadence to check for new AI if 24 hours is the wrong unit?

It depends on the layer. For frontier-model releases at the major labs, once a week is more than enough; major drops are announced on the lab's own blog and propagate within hours. For Hugging Face papers and code, daily skimming of huggingface.co/papers/trending catches what matters; sub-daily polling is mostly noise. For the agent and tooling layer (Claude Code, Codex, Cursor, Continue, MCP servers, computer-use harnesses), watching a small number of project changelogs on a rolling 7-day window beats any trending feed. The Fazm worked example for May 22, 2026 is that a single project's dated record, read across one day, often reveals more about ecosystem maturity than any aggregator does.

From this site