AI Research Roundup: Drug Discovery, Benchmarks, July 2026

Three distinct threads dominated AI research news this week: an OpenAI researcher is leaving to build a drug-discovery startup targeting a $2 billion valuation, a new benchmark called Singularity Gate is stress-testing frontier models on post-cutoff scientific prediction, and Thinking Machines released Inkling — a 975-billion-parameter open-weights model with third-party benchmark scores that rival closed competitors.

Miles Wang Exits OpenAI for $2B Drug Discovery Startup

Miles Wang, an OpenAI researcher whose published work includes evaluating how AI models can automate and accelerate scientific discovery, is leaving the company to found an AI drug-discovery startup, according to TechCrunch, which cited four people with knowledge of his plans. Wang joined OpenAI in 2024 after leaving Harvard, where he was pursuing a computer science degree. Several other OpenAI researchers are expected to join the new company.

According to TechCrunch, Wang is in talks to raise approximately $200 million at a $2 billion valuation, with Lightspeed in discussions to lead the round. Wang disputed the story’s funding figures and description of the company but did not provide corrected numbers. Lightspeed did not respond to a request for comment.

Two sources told TechCrunch the startup may focus on finding new uses for existing drugs — including those that previously failed in trials — rather than developing new compounds from scratch. Repurposing FDA-approved drugs can compress time to revenue significantly, since safety testing is already complete.

The funding discussions reflect a broader wave of investor interest in AI-driven life sciences. Chai Discovery, a two-year-old molecular interaction modeling startup co-founded by former OpenAI researcher Josh Meier, raised $400 million at a $3.8 billion valuation this week. Google DeepMind spinout Isomorphic Labs raised a $2.1 billion Series B in May.

Singularity Gate Benchmark Tests Post-Cutoff Scientific Prediction

A benchmark called Singularity Gate, designed to measure whether frontier AI models can predict paradigm-breaking scientific discoveries published after their training cutoff, published new results this week covering Fable 5 and GPT-5.6 Sol, according to a post on Reddit’s r/singularity.

The benchmark is notable for its framing: rather than testing recall of known facts, it probes whether models can extrapolate to discoveries they were never trained on — a harder and arguably more meaningful test of scientific reasoning capability.

According to the Reddit post, Claude Fable 5 currently leads the benchmark. The original Fable 5 responded to 45% of benchmark tasks, while the latest version responded to only 39%, with a slight performance degradation observed on the tasks both versions attempted. The post noted GPT-5.6 Sol represents a measurable improvement over GPT-5.5 on the same benchmark.

The refusal rate difference between Fable 5 versions — original versus latest — raises a recurring question in model evaluation: whether safety-tuning updates that reduce refusals on sensitive topics also affect performance on neutral scientific reasoning tasks.

Thinking Machines Releases Inkling, a 975B Open-Weights Model

Thinking Machines, the AI startup founded by former OpenAI CTO Mira Murati, released Inkling this week under an Apache 2.0 license — its first major open-weights model, available immediately on Hugging Face and the company’s own Tinker API, according to VentureBeat.

Inkling is a natively multimodal Mixture-of-Experts system with 975 billion total parameters, capable of processing text, images, and audio. A lighter variant, Inkling-Small at 276 billion parameters, was also announced in preview for cost-sensitive workloads.

Benchmark Performance

On third-party evaluations reported by VentureBeat, Inkling posted:

• 77.6% on SWE-bench Verified (software engineering), beating NVIDIA Nemotron 3’s 71.9%
• 91.4% on VoiceBench (voice understanding), behind Gemini 3.1 Pro’s 94.4% on high reasoning effort

Thinking Machines also said Inkling was designed to “answer directly on topics that may be subject to censorship” — positioning it for enterprises that prioritize factual outputs over content moderation defaults.

John Schulman, a Thinking Machines co-founder, shared the announcement on X. Horace He, a researcher at Thinking Machines previously from PyTorch, noted the engineering complexity involved, writing that “it truly takes a village” to ship a model at this scale.

PsiQuantum’s Photonic Quantum Computer Targets Drug Simulation

PsiQuantum, founded in 2016 by four UK physicists, is building a photonic quantum computer it claims could eventually compress decade-long drug simulation timelines, according to MIT Technology Review. The machine does not yet exist in operational form, but the company’s published architecture calls for roughly 100 stainless-steel cabinets cooled to near absolute zero using liquid helium, housing chips through which individual photons travel through optical switches and beam splitters.

Philipp Ernst, vice president at PsiQuantum, told MIT Technology Review that estimating how cytochrome P450 enzymes interact with specific drug molecules — a key step in pharmaceutical design — can take over 10 years with current computational methods. The company’s argument is that a sufficiently large photonic quantum computer could reduce that timeline substantially, though no operational timeline or performance figures were provided in the published account.

PsiQuantum operates in a field where even leading quantum prototypes remain too small and error-prone for practical use. The company’s photonic approach — using particles of light rather than superconducting qubits — is one of several competing architectures, none of which has yet demonstrated fault-tolerant computation at scale.

What This Means

The week’s research news, taken together, points to a maturing but still fragmented AI science ecosystem. On the application side, the clustering of former OpenAI researchers into drug-discovery ventures — Wang’s new startup, Chai Discovery’s Meier — suggests that life sciences is becoming a primary destination for frontier AI talent, backed by nine- and ten-figure funding rounds.

On the model side, Inkling’s release under Apache 2.0 adds a high-parameter open-weights option with competitive benchmark scores, narrowing the gap between open and closed models on specific tasks like software engineering. The censorship-resistance positioning is a deliberate enterprise differentiator, though how it performs in practice on sensitive real-world queries remains to be tested independently.

The Singularity Gate benchmark deserves attention as a methodology, not just a leaderboard. Testing whether models can predict post-cutoff discoveries is a more rigorous proxy for scientific reasoning than standard knowledge-recall benchmarks — and the observed performance drop between Fable 5 versions on refusal rate, without a corresponding explanation from the model developer, is the kind of regression that warrants scrutiny.

FAQ

What is Miles Wang’s AI drug discovery startup working on?

According to TechCrunch, the startup may focus on finding new therapeutic uses for existing or previously failed drugs rather than developing new compounds. Wang disputed the publication’s funding figures but did not clarify the company’s focus.

What does the Singularity Gate benchmark actually measure?

Singularity Gate tests whether frontier AI models can predict scientific discoveries published after their training cutoff — meaning it evaluates forward extrapolation rather than memorized knowledge. The benchmark’s Reddit post reported Fable 5 currently leads, though its latest version shows a lower response rate than the original.

How does Thinking Machines’ Inkling compare to other open-weights models?

Inkling scored 77.6% on SWE-bench Verified, outperforming NVIDIA Nemotron 3’s 71.9% on software engineering tasks, according to VentureBeat. On VoiceBench it scored 91.4%, trailing Gemini 3.1 Pro at 94.4%, making it competitive but not the top performer across all modalities.