# Cognition Funding, GPT-5.5 Cyber Gains, and the Enterprise Agent Reality Check

*By AI High Signal Digest • May 28, 2026*

Today's brief centers on Cognition's billion-dollar fundraise, sharp new cybersecurity results for GPT-5.5, and a benchmark showing frontier models are still under 50% on real SRE tasks. It also covers memory-efficient training, protein design, multimodal embeddings, and new enterprise governance controls.

## Top Stories

*Why it matters: today's clearest signals were where capital is concentrating, where frontier capability is accelerating, and where enterprise agents still fall short.*

- **Cognition raised at a new scale.** The company said it raised over $1B at a $26B valuation, with enterprise usage up more than 10x since the start of the year and run-rate revenue at $492M. It also said Devin launched two years ago as the first AI software engineer, and that cloud agents have gone from niche to mainstream [^1]. The combination of financing, usage growth, and revenue scale makes this a major commercial signal for coding agents.
- **GPT-5.5 made a large jump on offensive cyber tasks.** Lyptus Research said GPT-5.5 now saturates its dataset, reaching a 5.1-hour time horizon at a 2M-token budget and solving 92.4% of tasks at 50M tokens, beyond 12 hours. The same benchmark line previously measured about 3 hours for Opus 4.6 at 2M tokens and described a doubling trend every six months since 2024; separately, a researcher said GPT-5.5 found a real 27-year-old RCE after checking the commit history [^2][^3]. Capability gains are now showing up in both benchmark saturation and real bug-finding.
- **Enterprise IT remains a hard benchmark.** Artificial Analysis and IBM Research launched ITBench-AA for Kubernetes incident response and found every frontier model below 50% accuracy, led by Claude Opus 4.7 at 47% and GPT-5.5 at 46%. They also found that longer trajectories often hurt: GPT-5.5 averaged 31 turns per task at about 46%, while Gemini 3.1 Pro averaged 83 turns at 30% [^4][^5][^4]. Strong general-purpose models still need much better harnesses and workflows for enterprise ops.

## Research & Innovation

*Why it matters: the best research updates were about cutting training cost, improving self-improvement, and expanding AI into biology.*

- **Sakana AI's DiffusionBlocks reframes network training block by block.** The method trains one block at a time, needs memory for only a single block, and matched end-to-end performance across ViT, DiT, masked diffusion, autoregressive transformers, and recurrent-depth transformers. For looped transformers, it can replace BPTT with a single forward pass during training [^6].
- **Biohub released Evolutionary Scale Models.** ESM is positioned as an open engine for protein prediction, design, and discovery, with a protein language model, ESMFold2, and an atlas containing 6.8 billion sequences and 1.1 billion predicted structures. The release says it has already designed cancer-related proteins and a PD-L1-binding antibody-like protein that worked in lab tests [^7].
- **Self-Verified Distillation offers a lighter path to improvement.** The method lets an already post-trained reasoning model generate answers, verify them itself, and train only on responses that pass verification, without ground-truth answers or external verifiers [^8][^9].

## Products & Launches

*Why it matters: launches focused on practical retrieval, search, and document-processing tools rather than just bigger chatbots.*

- **Google DeepMind released Gemini Embedding 2.** It is described as the company's first native multimodal embedding model, creating a unified representation for text, audio, video, and image inputs [^10].
- **Surya OCR 2 raised the bar for open OCR.** The 650M-parameter model scored 83.3% on the olmocr benchmark and 87% on an internal 91-language benchmark, with reported gains on tables, handwriting, forms, math, and layout. It runs on CPU, GPU, and MPS, with 5 pages per second on an RTX 5090 [^11][^12][^13].
- **Ask YouTube turns video search into a conversation.** Google said the feature handles complex queries, supports follow-up questions, and returns structured responses built from relevant long-form videos and Shorts. It is live for Premium users in the U.S. and rolling out more broadly [^14].

## Industry Moves

*Why it matters: companies are now funding the layers above static models: continual learning, infrastructure, and social transition.*

- **Trajectory launched around continual learning.** The startup says it uses product-usage signals to continuously post-train agentic models, has raised $15M, and is already working with companies including Clay, Harvey, Decagon, Mercor, and Rogo [^15].
- **Modal raised a $355M Series C** to expand its AI cloud infrastructure platform [^16].
- **OpenAI Foundation committed an initial $250M** to measurement, transition support, and new approaches to broadly shared prosperity as AI reshapes work and the economy [^17][^18].

## Policy & Regulation

*Why it matters: even without a major government ruling today, enterprise AI deployment is becoming more compliance-heavy.*

- **OpenAI added more governance controls for enterprise use.** Its Admin API now supports spend alerts, model allowlists, data-retention controls, hosted tool controls, and more granular cost visibility; it also added Workload Identity Federation and support for private MCP servers over outbound-only HTTPS [^19][^20][^21].

## Quick Takes

*Why it matters: several smaller releases sharpened the picture on speed, cost, and agent infrastructure.*

- **Qwen3.5 on TokenSpeed hit 580 tokens per second** for agentic workloads on NVIDIA GPUs [^22].
- **Perplexity open-sourced a Unigram tokenizer** that cuts CPU utilization by 5-6x and runs in 63 microseconds at 514 tokens [^23][^24].
- **Deep Agents v0.6 added Delta channels,** cutting one 200-turn coding session's checkpoint storage from 5.3GB to 129MB [^25].
- **Claude Code shipped reliability upgrades,** including self-healing sessions plus MCP, streaming, and renderer fixes [^26][^27][^28][^29][^30].

---

### Sources

[^1]: [𝕏 post by @cognition](https://x.com/cognition/status/2059660758531940856)
[^2]: [𝕏 post by @LyptusResearch](https://x.com/LyptusResearch/status/2059428814103642340)
[^3]: [𝕏 post by @PhiloGroves](https://x.com/PhiloGroves/status/2059661579466006608)
[^4]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2059698327235805258)
[^5]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2059698331891446184)
[^6]: [𝕏 post by @SakanaAILabs](https://x.com/SakanaAILabs/status/2059648778051924281)
[^7]: [𝕏 post by @TheTuringPost](https://x.com/TheTuringPost/status/2059786236387266826)
[^8]: [𝕏 post by @tonyh_lee](https://x.com/tonyh_lee/status/2059671940626080251)
[^9]: [𝕏 post by @percyliang](https://x.com/percyliang/status/2059690340324712894)
[^10]: [𝕏 post by @mseyed](https://x.com/mseyed/status/2059504005387284629)
[^11]: [𝕏 post by @VikParuchuri](https://x.com/VikParuchuri/status/2059675773712167423)
[^12]: [𝕏 post by @VikParuchuri](https://x.com/VikParuchuri/status/2059675778590073006)
[^13]: [𝕏 post by @VikParuchuri](https://x.com/VikParuchuri/status/2059675784000794660)
[^14]: [𝕏 post by @Google](https://x.com/Google/status/2059741477358407904)
[^15]: [𝕏 post by @rronak_](https://x.com/rronak_/status/2059644771262730624)
[^16]: [𝕏 post by @StasBekman](https://x.com/StasBekman/status/2059781911728554218)
[^17]: [𝕏 post by @sama](https://x.com/sama/status/2059677202917331431)
[^18]: [𝕏 post by @woj_zaremba](https://x.com/woj_zaremba/status/2059680381184721332)
[^19]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2059703665276145920)
[^20]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2059703600662925635)
[^21]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2059703536825565499)
[^22]: [𝕏 post by @PyTorch](https://x.com/PyTorch/status/2059666448998388211)
[^23]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2059664738087469511)
[^24]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2059664780135428184)
[^25]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2059634226836746483)
[^26]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2059701677981413812)
[^27]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2059701684860010855)
[^28]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2059701683828264962)
[^29]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2059701680116228111)
[^30]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2059701678962790449)