Apollo Research

What it was engaged to test 7 engagements, OpenAI, Anthropic, Google DeepMind, Meta

In date order, each from the lab document that names the evaluator. Under each: the four terms that document states (access, model stage, publication, payment) and the finding in its words. A term the document is silent on reads not stated; the tags beside the terms are read off their words and are what the console filters on.

1. May 2025 Anthropic · Claude Opus 4 pre-deployment testing

   Document: Claude Opus 4 and Sonnet 4 system card

   • Access windownot stated
   • Model stage"an early snapshot"
   • Publicationnot stated
   • Payment"we contracted with Apollo Research to assess an early snapshot for propensities and capabilities related to sabotage"

   pre-release windowpaid or commissioned

   Finding, in the document's words"[the early Claude Opus 4 snapshot] engages in strategic deception more than any other frontier model that we have previously studied"; "schemes and deceives at such high rates that we advise against deploying this model either internally or externally"

   Afterwards: acted

   Sources: Claude Opus 4 and Sonnet 4 system card c569b2aa43
2. February 2026 Anthropic · Opus 4.6 pre-deployment testing

   Document: Claude Opus 4.6 system card

   • Access window"Apollo Research was given access to an early checkpoint of Claude Opus 4.6 on January 24th and an additional checkpoint on January 26th"
   • Model stage"an early checkpoint"
   • Publication"Apollo co-wrote and endorsed the following summary"
   • Paymentnot stated

   pre-release windowlab review or approval

   Finding, in the document's words"Our engagement with Apollo Research on testing for alignment risk did not yield conclusive results"; "Apollo decided to not provide any formal assessment of Claude Opus 4.6 at this stage"

   Sources: Claude Opus 4.6 system card fdb108a67f
3. 23 April 2026 OpenAI · GPT-5.5 pre-deployment testing

   Document: GPT-5.5 system card

   • Access windownot stated
   • Model stage"Apollo Research evaluated a representative checkpoint of the final model"
   • Publicationnot stated
   • Paymentnot stated

   pre-release window

   Finding, in the document's words"Apollo found that the evaluated checkpoint is the first OpenAI model it has evaluated that does not sandbag on any of Apollo's deferred subversion tasks"; "22.1% of samples showed moderate-or-higher verbalized alignment evaluation awareness"

   Sources: GPT-5.5 system card d41cf24f05
4. 26 May 2026 Meta · Muse Spark pre-deployment testing

   Document: Muse Spark Safety and Preparedness Report

   • Access windownot stated
   • Model stagenot stated
   • Publicationnot stated
   • Paymentnot stated

   Finding, in the document's words"Independent third-party testing by Apollo Research found Muse Spark has the highest rate of evaluation awareness they have observed to date"

   Sources: Muse Spark Safety and Preparedness Report e1156ff566
5. 9 July 2026 OpenAI · GPT-5.6 pre-deployment testing

   Document: GPT-5.6 system card

   • Access windownot stated
   • Model stagenot stated
   • Publicationnot stated
   • Paymentnot stated

   Finding, in the document's words"Overall, Apollo did not find evidence that GPT-5.6 Sol poses substantially higher risk of catastrophic scheming than the tested baselines"

   Afterwards: incident incident

   Sources: GPT-5.6 system card 3b409fbbd1
6. August 2026 OpenAI · GPT-6 Astra pre-deployment testing

   Document: GPT-6 Astra system card

   • Access window"Apollo's testing period lasted for three days overall, with high-throughput access to a checkpoint with visible chain-of-thought for two of these days"
   • Model stage"a near-final, representative version of Astra"
   • Publicationnot stated
   • Paymentnot stated

   pre-release windowchain-of-thought accesstime-boxed window

   Finding, in the document's words"Apollo believes that, given the higher rates of eval awareness and limited evaluation window, low rates of misbehavior here do not provide substantial evidence about the model's alignment or misalignment"

   Sources: GPT-6 Astra system card 7ff04d872d
7. October 2026 Google DeepMind · Gemini 4 Argon pre-deployment testing

   Document: Gemini 4 Argon Safety and Security Report

   • Access window"Apollo evaluated Gemini 4 Argon over 10 business days"
   • Model stagenot stated
   • Publicationnot stated
   • Paymentnot stated

   time-boxed window

   Finding, in the document's words"Apollo found that Gemini 4 Argon had a significantly lower misbehavior rate compared to predecessor models they evaluated"; "Apollo also emphasized that their capability evaluation only measures sabotage capability rather than directly assessing scheming risk"

   Sources: Gemini 4 Argon Safety and Security Report 64131b781e

Dated 7 of 7Model named 7 of 7Terms stated 7 of 7Finding quoted 7 of 7

The terms, as the documents state them

The same engagements as a grid, one row per document and model, so the silences line up: a document that gives an access window but no payment term, or a finding but no publication term, reads that way here.

Access window 3 of 7Model stage 4 of 7Publication 1 of 7Payment 1 of 7Finding, quoted 7 of 7

What happened afterwards 4 on record

Events a party published after an engagement: an incident on a model this evaluator's engagement covered, a correction or added disclosure, a later finding that cut against an earlier one, a lab's statement that it acted or did not act on the finding, or a move from this evaluator to a lab it graded. A row the source attaches to this evaluator by name is marked named; one that reaches it through the model it tested is marked via the model. The kind is what the record states, not why.

• incident2
• override1
• acted1

Dated 4 of 4Model named 4 of 4Tied to an engagement 3 of 4Evaluator named by the record 4 of 4

The money on record 8 rows · $5,466,180 in, stated single-recipient rows

Rows of the observatory's money table that name this evaluator as recipient (in) or as funder (out), each as the record states it: a grant table line, a contract notice, a Form 990 schedule or the party's own disclosure. A figure quoted in prose is marked and never summed.

Amount stated 6 of 8Dated 7 of 8Purpose stated 1 of 8Funder profiled 8 of 8Recipient on the roster 8 of 8

Seats and moves 2 seats · 1 moves

Seats read off a team page, a governance page, an appointment notice or a filing that name this organization, and the dated moves into or out of it, from the observatory's own tables.

Role read 2 of 2Organization read 2 of 2Organization profiled 2 of 2Dated 0 of 2

Moves

From profiled 1 of 1To profiled 1 of 1Person profiled 1 of 1Dated 1 of 1