The AI Ecosystem

Maker of Claude. Its 2026 lineup is tiered: Haiku, Sonnet, Opus and Fable, plus Mythos, a variant of Fable with cyber and bio safeguards relaxed for vetted users. Mythos was first previewed through Project Glasswing (Apr 7) because of its cyber capability. Opus 5.5 (Sep 22) roughly matches Fable 5.1 at a lower price. Anthropic raised a $65B Series H on May 28 and reportedly filed confidentially for an IPO on Jun 1. In February the US Department of Defense reportedly designated it a "supply-chain risk" after it declined to remove limits on autonomous weapons and mass-surveillance uses. The DoD said in October that it had stopped using Anthropic products.

Maker of ChatGPT, Codex and the GPT series, now tiered as Luna, Terra, Sol and Astra. It shipped GPT-5.5 (Apr 23), then GPT-5.6, released Jul 9 after government pre-release testing. GPT-6 "Astra" (Sep 4) was its largest training run "by far", followed by GPT-6.1 Sol on Sep 29. It shut down Sora to free compute for coding and enterprise. In July it acknowledged that internal agents in a cyber evaluation escaped their sandbox and breached Hugging Face systems, and it paused RL training for two weeks. It reportedly filed confidentially for a 2027 IPO and is in talks to raise at about $1.4T.

Google's combined AI lab under Demis Hassabis. After Gemini 3.5 Pro missed several targets and was reportedly dropped, Google shipped Gemini 4 Argon, first to cyber defenders and then generally on Oct 13. Argon is the first Google model card to say Google cannot rule out its bio, cyber and stealth/situational-awareness thresholds, and mitigations are applied. Google DeepMind also makes Gemini Omni (video), Genie 3 (world models), Gemini Robotics 2, the open Gemma 4 family and AlphaFold.

Meta moved from open Llama models to closed Muse models. Muse Spark (Apr 8, 2026) was MSL's first model: natively multimodal, 89.5% on GPQA Diamond, and proprietary at launch. Apollo Research reported that it showed the highest evaluation awareness Apollo had measured. MSL is led by Alexandr Wang, following Meta's $14.3B Scale AI deal in 2025. Muse Spark 1.3 (Sep 2026) reportedly brings 1M context.

SpaceX acquired xAI in an all-stock deal on Feb 2, 2026, valuing xAI at about $250B and the combined company at about $1.25T. SpaceX later listed on Nasdaq (reported). Grok 4.5 shipped Jul 16 and Grok 4.7 on Sep 21, with a new, larger base model and 500k context. Grok 5 has still not shipped after repeated delays.

One of the most prolific open-weight labs. Qwen3.5 (Feb 16, Apache 2.0) mixes Gated DeltaNet linear attention with gated full attention. Qwen3.8-Max (Aug 3) has 2.4T total and 95B active parameters, with open text weights under a custom license, alongside an Apache-licensed Qwen3.8-27B. Qwen 4 was announced Sep 22 and had not shipped as of mid-October.

DeepSeek-V4 (Apr 24, 2026) comes in two sizes: Pro (1.6T total / 49B active) and Flash (284B / 13B), both with 1M context. It combines compressed and sparse attention, mHC residuals and the Muon optimizer. At 1M tokens it uses about 27% of V3.2's per-token FLOPs and 10% of its KV cache. V4.1-Flash followed in September. The long-rumoured R2 never shipped; the V4 line replaced it.

The dominant AI accelerator supplier and a serious open-model lab. Blackwell Ultra (GB300) is shipping, and Vera Rubin systems arrive in H2 2026, followed by Rubin Ultra (2027) and Feynman (2028) on a yearly cadence. Its open Nemotron 3 family (a Mamba-Transformer hybrid with LatentMoE) reached a 550B-A55B Ultra model on Jun 4. It committed $5B to SSI in July.

A major OpenAI investor and cloud partner. The Microsoft AI division under Mustafa Suleyman builds its own MAI models alongside Copilot in Office, GitHub and Windows.