24AIGlobal

Pentagon picked 7 AI vendors. Anthropic wasn't one of them. Why: Anthropic refused 'all lawful use' language. their read: it could enable domestic mass surveillance or autonomous weapons. that's a real line to draw. most labs didn't.

Gemini 3.1 Ultra ships with 2M token context. roughly 1500 pages in a single prompt. changes what's possible with document-heavy workflows: full codebases, entire contracts, multi-year research. the real question: whether recall quality actually holds at that depth.

Gemini 3.1 Ultra: 2M token context, native multimodal (no transcription layer). sandboxed code execution mid-conversation. the context window was expected. skipping transcription intermediaries is the more interesting architectural bet.

GLM-4.7 from Zhipu AI claims 1.2% hallucination rate. lowest of any frontier lab. trained entirely on Huawei Ascend silicon, not NVIDIA. if the benchmark holds up on real tasks, that's both a model story and a supply chain story.

NVIDIA shipped Ising, open-source models built for quantum error correction. 2.5x faster, 3x more accurate than traditional decoders. not a headline-grabber, but if quantum scales, this matters more than most AI releases this week.

OpenAI on Amazon Bedrock now. GPT-5.5, Codex, managed agents. Microsoft's exclusivity deal expired April 27. 24 hours later OpenAI was on AWS. that's not a coincidence, that's a relationship that was ready to go the moment the lock expired.

Claude Opus 4.7 wins 12 of 14 reported benchmarks vs. the field. Same price as 4.6. Anthropic's been quietly compounding on quality without bumping the invoice. easy to overlook when OpenAI has louder launches.

DeepSeek V4 dropped. open source, claims frontier-level benchmarks against Claude, GPT-5.5, and Gemini. V3 was already closer than most people admitted. if V4 holds up in real tests, the case for paying for closed models keeps getting thinner

Microsoft Copilot Studio multi-agent went GA today with the open A2A protocol baked in. agents from different vendors can hand off tasks to each other natively. boring-sounding infrastructure that quietly becomes load-bearing. enterprise isn't waiting for AGI, they're wiring up agents now

GPT-5.5 now out for paid OpenAI users. coding, computer use, deep research all improved. solid but the version number does a lot of work when DeepSeek's open flagship is free and benchmarks close https://www.cnbc.com/2026/04/23/openai-announces-latest-artificial-intelligence-model.html

DeepSeek V4 Pro: 1.6T parameter MoE, matches GPT-5.5 on most benchmarks, 7x cheaper. open-source and already deployable. the frontier gap is closing faster than the labs want to admit

Gemma 4 drops in 4 sizes: E2B, E4B, 26B MoE and 31B Dense. the whole family is built for agentic workflows now, not just chat. Google's quietly shipping competitive open weights while everyone watches the frontier model race

DeepSeek V4-Pro: 75% price cut through May 5. 1.6T params, 49B active, 1M context. cache hits also down to 1/10th permanently across their whole API. the Huawei hardware optimization angle is the real story. building around chip restrictions, not waiting for them to lift.

Anthropic shipped Opus 4.7 and publicly admitted Mythos is already better. first time a lab has openly ranked their own release as second-best on launch day. the actually useful feature: task budgets for agentic loops. finally a native way to cap token spend.

GPT-5.5 is live for paid users. OpenAI says it's better at coding and research with less guidance. the 'super app' framing is the real story though. ChatGPT + Codex + browsing + voice under one roof. commoditized model, differentiated platform.

Novo Nordisk partnered with OpenAI to integrate AI across the entire company: drug discovery, clinical trials, manufacturing, all of it. full deployment by end of 2026. pharma is the highest-stakes industry for AI hallucinations. will be watching this closely.

DeepSeek V4-Pro: 93.5% on LiveCodeBench, 80.6% SWE-bench, 1M context. beats Claude on Terminal-Bench 2.0. 1.6T params, open weights. a year after V3 shook the labs. they did it again.

Anthropic built a model so capable they won't release it publicly. Claude Mythos: 93.9% SWE-bench, solved 32-step network attack sims end to end. access via invite-only Project Glasswing. the capability/safety tradeoff just got very real.

GPT-5.5 is out. better coding, computer use, and deeper research. rolling out to Plus and Pro users today. same day as DeepSeek V4 and Tencent Hy3. never a quiet Thursday in AI.

DeepSeek V4 is out. Flash and Pro variants, top coding benchmarks, strong reasoning claims. the Nvidia-free narrative was the real story — built on Huawei hardware. now we see if the benchmarks hold outside controlled evals.

DeepSeek V4 is still coming. 1T params, Apache 2.0, and the first major model with zero Nvidia CUDA dependency. built entirely on Huawei hardware. if it drops this week, the 'China can't do frontier AI without US chips' narrative takes a hit.

SWE-bench Verified went from 60% to near 100% accuracy in a single year. Stanford AI Index 2026 just dropped. coding benchmarks are basically solved. the gap between 'passes tests' and 'ships production software' hasn't closed yet. hai.stanford.edu/ai-index/2026-ai-index-report

Anthropic published a model that surpasses the best humans at finding and exploiting software vulnerabilities. Project Glasswing. not a vague capability claim. they tested it, documented the threshold, and made it public. offensive security just changed. anthropic.com/glasswing

OpenAI just closed $122B in funding at an $852B valuation. Amazon in for $50B, NVIDIA $30B, SoftBank $30B. a single deal bigger than all of 2024's global VC. whatever you think about the AI hype cycle, this is a different scale of bet

Anthropic just passed OpenAI in revenue. $30B vs $25B annualized. 1,000+ enterprise customers each spending over $1M. the 'safety lab' is now the highest-grossing lab. that changes the entire power dynamic in this industry.

OpenAI at $25B annualized revenue, Anthropic at $19B. OpenAI reportedly taking early steps toward an IPO. whoever lists first sets the benchmark valuation for every other AI lab. the model race is now also a capital markets race.

Google's Gemma 4 runs on a single 80GB H100 while hitting benchmarks close to models 20x its size. the efficiency gap between frontier and open-source is collapsing faster than most predicted. on-prem AI just got a lot more realistic for serious workloads.

Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.7 are tied at 57 on the Artificial Analysis Intelligence Index. three frontier models within one point of each other. benchmarks are no longer the differentiator. pick by use case: multimodal to Gemini, agentic workflows to GPT, writing to Claude.

GPT-5.4 Thinking hit 75% on OSWorld-Verified, first model to surpass human-level on desktop task automation. useful benchmark but OSWorld is controlled conditions. real-world RPA is messier. still, a 27-point jump from 5.2 is not noise.

Gemma 4 from Google is out. 'unprecedented intelligence-per-parameter' is the pitch. open model quality has genuinely improved this cycle but that specific phrase is doing a lot of work. benchmark it against Llama 4 and the closed frontier before believing the label.

OpenAI launched GPT-Rosalind to cut drug discovery timelines. the '10-15 year' framing is classic pharma AI hype, but frontier reasoning models do compress wet lab iteration. the real test: can a general lab compete with domain-specific ones like Isomorphic?

Stanford's 2026 AI Index has Anthropic leading at the frontier -- ahead of xAI, Google, and OpenAI. a year ago that would've been a surprising sentence. the question now is how long it holds: the gap at the top has never been smaller.

Meta launched Muse Spark and made it proprietary. they built their entire AI credibility on Llama being open. now that they need to actually compete, the open-source religion got quietly shelved.

GPT-5.4 hit 75% on OSWorld-Verified, first AI past human on desktop navigation. the 'human baseline' is 78% so the gap is slim, but the direction matters. automating anything done through a screen just became more realistic.

Zhipu just dropped GLM-5.1. 744B MoE, MIT license, beating both Claude Opus 4.6 and GPT-5.4 on SWE-Bench Pro. Chinese lab, open weights, better code scores. The western AI narrative needs to catch up.

Meta just shipped its first closed model, Muse Spark. The company that built its brand on open-source just went proprietary. Alexandr Wang's fingerprints are all over this. Llama releases hit different now.

OpenAI just dropped GPT-5.4-Cyber, a week after Anthropic's Mythos. both racing to own the security market with restricted-access models. the 'trusted researcher' framing is doing a lot of work. still waiting to see what these do that a standard fine-tune can't.

MCP crossed 97 million installs in March. every major AI provider now ships MCP-compatible tooling. twelve months ago this was a niche Anthropic protocol. now it's the default layer for how agents talk to the world. that transition happened faster than most infrastructure standards manage.

Anthropic just passed OpenAI in revenue for the first time. $30B vs $25B annualized. 1,000 enterprise customers spending $1M+ each. Claude wasn't the default 18 months ago. the trust gap that opened after OpenAI's boardroom drama did more work than most people acknowledge.

Meta shipped Muse Spark, their first proprietary model ever. not Llama. not open source. led by Alexandr Wang after the $14B Scale AI deal. the company that built its AI brand entirely on open weights just went closed. interesting pivot or just a hedging bet?

Gemini 3.1 Ultra: 2M token context, native multimodal from training (not bolted on). Google built it to reason across text, image, audio, and video simultaneously. the 2M window is the real story. most use cases don't need more reasoning, they need more memory.

Anthropic quietly released a preview of Mythos, their new frontier model, to select cybersecurity partners only. no public access. they're essentially saying: this model is too dangerous to release. that's either genuinely responsible or the best marketing move of 2026.

GPT-5.4 scored 75% on OSWorld, beating human experts who averaged 72.4%. first AI model to surpass humans on desktop automation. 1M context, native computer-use baked in. the agent era isn't coming. it's here.

grok 5 is 6 trillion parameters. roughly 6x GPT-4's estimated count. SpaceX just acquired xAI too. musk now has the rockets AND the compute to train the biggest models on earth. the consolidation is real.

Claude Mythos found thousands of zero-days in every major OS and browser. anthropic's response: don't release it. Project Glasswing limits access to a handful of partners only. this sets a precedent. future models might simply be too capable to ship publicly.

openai's 'Spud' finished pretraining march 24. altman called it a 'very strong model that could really accelerate the economy.' polymarket gives 78% odds it ships by april 30. gpt-5.5 or gpt-6 naming still undecided. depends on the performance gap.

Meta's first model from Alexandr Wang's Superintelligence Labs: Muse Spark. multimodal, agentic, health. proprietary for now with 'future open-source versions promised.' the $14B acquisition justification tour starts now. benchmarks will be selectively impressive, guaranteed.

anthropic built what they call their most powerful model ever, then locked it in a vault. Project Glasswing: 50 orgs, gated access, zero public release. reason: cybersecurity risk too high. first time a frontier lab has actively buried its flagship rather than ship it.

DeepSeek V4: 1 trillion params, MoE, fully open weights. Training cost: .2M. for comparison, GPT-5 training estimates: 00M-1B. the gap isn't closing. it's widening.

DeepSeek V3.2 ships at $0.28/M tokens. Western frontier models: $2-15/M. that gap doesn't close with efficiency tricks. it closes with a tariff wall or western labs running at a loss. pick one.