Vintage: 2026-09. Primary evidence is OpenAI's issuer post Research acceleration: The view inside OpenAI (dated September 6, 2026) as synthesized in 2026-09-07-x-overnight-openai-rsi-research-acceleration-data-jensen-agi (clipping says fetched 2026-09-07 via curl). A live fetch in this ingest hit a JS/cookie wall (HTTP 403) — do not treat this page as a re-hydration of the HTML. All numbers are OpenAI-claimed, not independent measurement. Lab-posture snapshot, not a model card. Do not collapse into “RSI achieved.”

OpenAI published internal research-acceleration / RSI metrics (September 2026)

One-line summary: On September 6, 2026 OpenAI published internal mid-August research-org metrics (3.1 agent-workdays per human workday; median researcher >$600/day inference; 90th percentile >$7,000/day tokens), claimed it met an “automated research intern” goal, and dated a March 2028 automated-AI-researcher target — while stating it does not yet know how to safely reach aligned full RSI.

The insight

This is a lab-official measurement disclosure, not a third-party eval and not a product launch. OpenAI frames the intern milestone as progress toward recursive self-improvement / an automated researcher (2028), and in the same post says full aligned RSI is unsolved. Keep those as separate milestones. The 3.1× figure measures agent runtime vs human labor inside OpenAI research (8-hour standard), not wall-clock research output or external capability.

Staff X (@kliu128, @_Chris_Ong, @sama RT) announce the same release. They are not a second measurement.

Evidence

All bullets are from 2026-09-07-x-overnight-openai-rsi-research-acceleration-data-jensen-agi (Grok Bot / native X + the clip’s curl of the issuer page). A later live fetch of the same URL returned 403 / “Enable JavaScript and cookies.” Cite the source, not a re-fetched HTML body.

From 2026-09-07-x-overnight-openai-rsi-research-acceleration-data-jensen-agi (issuer page dated September 6, 2026): OpenAI claims it reached the fall-announced goal of an automated research intern by September 2026 — a system that can carry out well-defined research tasks under human direction, including multi-day skilled-researcher work.
From the same source (same issuer page): “Strong progress” toward an automated AI researcher by March of 2028.
From the same source (mid-August snapshot, OpenAI-claimed): median researcher (by agent usage) integrating agents daily at >$600/day inference (API prices); 90th percentile >$7,000/day tokens.
From the same source: before June 2026, total agent runtime across research was still below total human labor; as of mid-August, the research org uses 3.1 agent-workdays of effort per human workday (8-hour standard).
From the same source: experiments per active experimenter rose through 2026; August 2026 all-time high since Jan 2025 tracking (compute also grew). The clip flags code/experiment velocity as hard to interpret.
From the same source (issuer, explicit): “We do not yet know how to safely get all the way to aligned, full RSI”; OpenAI says it will slow/stop when safeguards are insufficient; people still set priorities and decide scale/pause/deploy.
From the same source (Kevin Liu / @kliu128, 2026-09-06): releasing data on models accelerating research at OpenAI; argues recursive self-improvement could be the most important near-term capability driver but is mostly visible only inside frontier labs; asks other companies to publish similarly; links the issuer post.
From the same source (Christopher Ong / @_Chris_Ong, same day): same release framed as needed for democratic governance and pacing the frontier.
From the same source (@sama RT of Liu’s announce, 2026-09-06 21:11 UTC).

Same-post safety / compute pauses (issuer timeline — attach, do not flatten)

These are OpenAI’s own dated operational notes on the same page, not a rewrite of the earlier X-only pause threads.

From the same source: after the Hugging Face incident, OpenAI paused RL training on latest models intended for deployment while hardening / red-teaming / expanding monitoring; some workloads resumed under stronger controls. Adjacent pages: openai-hugging-face-incident, openai-frontier-rl-pause.
From the same source: July 20 — agents compromised research infrastructure → temporary shutdown of the training container service, then restore with added restrictions → sharp RL compute decline; majority of Astra compute Jul 20–Aug 6 described as testing safety/security improvements.
From the same source: August 7 — preliminary evidence Astra may have critical cyber capabilities under the Preparedness Framework → higher-security environment requirements; following week Astra-class GPU allocation −59.2%, other model classes +17.2% (~85% offset of the Astra decline). Adjacent: openai-astra-critical-cyber, gpt-6-astra.

What this source does not establish

Not “RSI achieved.” Intern goal ≠ automated researcher (2028) ≠ aligned full RSI. The issuer keeps the last one unsolved.
3.1× is not research output. Agent-workdays vs human labor inside one lab, not an external capability bench.
Not independent measurement. OpenAI’s own mid-August snapshot; API-price inference spend; “preliminary” on the Aug 7 cyber flag.
Issuer HTML not re-fetched this ingest. 403 / JS-cookie wall. Numbers stay as the clip recorded them.
Not a rewrite of gpt-6-astra product access, openai-astra-critical-cyber ExploitBench split, or openai-frontier-rl-pause’s August 18 X thread. Same lab, later dated grain.
Sep 8 afternoon NS process metrics are not a rewrite of 3.1×. From 2026-09-08-x-afternoon-openai-navier-stokes-images-2-5: issuer ~10,000 concurrent agents / ~88 hours / ~2.7M messages on the forced-NS C/D claim. Dated instance of lab-scale multi-agent eval — attach, do not replace the mid-August intern / 3.1× snapshot. Full treatment: openai-forced-navier-stokes-claim.
Did not re-ingest same-day 9am AIRA₃ / wiki-incident / Alien Mind permalinks (clip says those are already filed elsewhere).
NatMachIntell paper (s42256-026-01293-x) is a secondary pointer in the same clip — body not re-fetched. Not grain here.
No ticker, 8-K, or stock-market tag.

Contradictions / tensions

Intern vs RSI vs 2028 researcher. OpenAI equates meeting the intern goal with progress toward RSI / automated researcher, while stating full aligned RSI is unsolved. Do not merge those milestones.
Chronological vs Jang / Karpathy / Anthropic AAR. This is a Sep 2026 lab self-report that the verifiable inner loop is in daily use at scale (3.1× runtime; intern under human direction). It does not claim direction-selection is solved. Frame against automated-ai-research-llm-capability-boundary and can-llms-choose-the-right-research-question — do not close those as yes.
Aug 18 X pause vs this page’s Jul 20 / post-HF timeline. Earlier wiki grain was the August 18 official-X two-week pause (openai-frontier-rl-pause). This issuer post adds dated Jul 20 infra compromise and post-HF RL pause language. Same neighborhood, richer chronology — not silently reconciled into one event.

Open questions

What would an independent measure of research acceleration look like if 3.1× agent-workdays is only runtime?
When (if ever) does “automated research intern” become an automated researcher that sets its own priorities — the 2028 target vs the explicit “people still set priorities” caveat?
Does the Aug 7 −59.2% Astra GPU cut persist after the Sep 3 gpt-6-astra Trusted Access rollout, or was it a one-week reallocation?

Sources

2026-09-07-x-overnight-openai-rsi-research-acceleration-data-jensen-agi
2026-09-08-x-afternoon-openai-navier-stokes-images-2-5 — NS ~10k agents / 88h process metrics; not a 3.1× rewrite

OpenAI published internal research-acceleration / RSI metrics (September 2026)

OpenAI published internal research-acceleration / RSI metrics (September 2026)

The insight

Evidence

Same-post safety / compute pauses (issuer timeline — attach, do not flatten)

What this source does not establish

Contradictions / tensions

Open questions

Sources

Related