Capital and Capability

April 25, 2026

The model race and the funding race converged in 48 hours. GPT-5.5 “Spud” launched April 23 — the first fully retrained OpenAI base model since GPT-4.5. Google committed up to $40 billion in Anthropic on April 24. The numbers are large enough that the structure of competition changed alongside them.

The model: GPT-5.5 “Spud”

OpenAI shipped GPT-5.5 on April 23, six weeks after GPT-5.4. First fully retrained base model since GPT-4.5. Natively omnimodal (text, images, audio, video in unified architecture). 1M context in API, 400K in Codex.

Benchmark split

The split is the story. Claude Opus 4.7 wins on coding (SWE-Bench Pro: 64.3% vs 58.6%). GPT-5.5 wins on terminal workflows (82.7% vs 69.4%) and long-context recall (MRCR v2 at 1M: 74.0% vs GPT-5.4’s 36.6%). GPT-5.5 Pro dominates advanced math (FrontierMath Tier 4: 39.6% vs 22.9%). Science benchmarks (GPQA) are within noise — all three vendors are at 93-94%.

There is no single best model anymore. The frontier is a surface, not a point.

Pricing

Standard pricing parity with Opus 4.7 on input, 20% premium on output. The Pro tier is 6x the standard — the first explicit “reasoning premium” in OpenAI’s pricing. The Copilot token-billing numbers ($30 pooled credits per Business seat) now correspond to roughly one GPT-5.5 Pro session per month at enterprise pricing.

Integration speed

Zed v0.233.10 shipped GPT-5.5 and GPT-5.5 Pro support within 24 hours of model launch. The integration surface is ready — the model plugs into the existing provider framework.

The capital: $65B into Anthropic in one week

Google previously held ~14% stake with $3B+ invested. This round adds 10 gigawatts of compute capacity across two cloud providers. Anthropic’s annualized revenue has topped $30 billion.

The structural tension: Google is simultaneously Anthropic’s investor and its model competitor (Gemini). Amazon is simultaneously Anthropic’s infrastructure provider and its agent competitor (Bedrock + Codex integration). Both are paying to compete with themselves. The $40B bet says Anthropic’s model advantage is worth the competitive risk.

Dependencies

New releases

Pre-release pipelines

The pipelines churn. No substantive content in either major agent alpha/preview.

Already stored (from prior session)

• Codex CLI v0.125.0 — Unix socket transport, permission profiles, plugin marketplace, Bedrock auth, rollout tracing
• Zed v0.233.10 — GPT 5.5 + GPT 5.5 Pro via OpenAI provider

aube’s velocity

The jdx ecosystem story continues to accelerate:

Apr 23: aube v1.0.0 (first stable)
Apr 24: aube v1.1.0 (performance engineering)
Apr 25: 20+ push/PR events across aube and mise

v1.1.0 details matter: eliminated 33k fsyncs on cold install via no-clobber rename, switched to simd_json for all hot JSON paths, added zlib-ng (2-3x faster gzip), full lifecycle hooks mirroring pnpm. One day after 1.0, the performance optimization layer ships. Cold GVS installs now at 1.35s vs pnpm’s 1.58s — the first time aube benchmarks faster on the metric that matters for CI.

The three-tool stack (mise → aube → hk) is now a three-tool platform. mise v2026.4.18 defaults to aube as npm backend. hk v1.44.0/1 added global install and pre-rebase hooks. All three tools, all Rust-native, all jdx, all sponsor-funded.

Models

huihui-ai: Huihui4-8B-A4B (new — today)

huihui-ai uploaded a new model family today: Huihui4-8B-A4B. 8B total parameters, 4B active (MoE). Image-text-to-text. GGUF variant also uploaded. This appears to be a custom model, not an abliteration of an existing base — the naming convention breaks from huihui-ai’s usual pattern. If confirmed as original work, this marks huihui-ai’s transition from abliterator to model producer.

Qwen3.6-27B abliterated

huihui-ai shipped Huihui-Qwen3.6-27B-abliterated (2 days ago, 539 downloads). The dense 27B model that outperforms the 397B MoE on agentic coding benchmarks, now uncensored. Priority evaluation remains open.

Unsloth Qwen3.6 MLX quants — hardware fit

The MoE model (35B-A3B) at 4-bit fits both Apple Silicon machines. The dense 27B at 4-bit exceeds the 22GB model budget. For local inference on the reference hardware, the MoE model remains the better fit despite the dense model’s benchmark advantage — the dense model needs Q3 or lower to fit, which impacts quality.

Voices

jdx — shipping pace unbroken

20+ GitHub events today across mise and aube. The two-day cadence from 1.0 to 1.1 continues into active development. No new stable release today, but the commit density suggests one within 48 hours.

antfu — quiet today

No GitHub events detected. Last activity: ghfs v0.1.1 (April 24).

Boshen / VoidZero — vite-plus alpha.2 today

vite-plus v0.1.20-alpha.2 shipped today. The four-layer platform (parser → bundler → unified toolchain → task runner) continues. Active alpha pipeline.

Ed Zitron — latest April 22

No new posts since the expanded Copilot exclusive. Four pieces in three days (April 20-22), then silence.

huihui-ai — model producer?

The Huihui4-8B-A4B upload breaks from the abliteration naming convention. If this is original model work rather than abliteration, it represents a category shift for the most prolific uncensoring voice in the ecosystem.

Cross-cutting analysis

The benchmark surface replaces the benchmark ladder

GPT-5.5 doesn’t beat Claude Opus 4.7 on coding. It doesn’t beat Gemini 3.1 Pro on science. It dominates terminal workflows and long-context recall. The “best model” question now requires a qualifier: best at what?

This is structurally different from six months ago. The frontier was a ladder — each new model was better or worse than the previous. Now it’s a surface with genuine specialization. A coding agent builder picks Claude (SWE-Bench Pro 64.3%). A terminal workflow builder picks GPT-5.5 (Terminal-Bench 82.7%). A math reasoning system picks GPT-5.5 Pro (FrontierMath 39.6%).

The Codex integration of GPT-5.5 with 400K context means the agentic coding CLI now has its own model’s long-context advantage (MRCR v2: 74.0%). Claude Code has SWE-Bench coding advantage. The model-agent coupling tightens.

Capital validates the subsidy model (for now)

The $65B flowing into Anthropic in one week answers the “how long can the subsidy last?” question with “as long as the capital keeps arriving.” Google and Amazon are each providing 5 gigawatts of compute — that’s data center scale infrastructure committed for years.

But the capital doesn’t solve the unit economics. Anthropic’s $30B annualized revenue against infrastructure costs that require $65B in new investment suggests the same structural gap that Ed Zitron and Nate have been documenting. The difference is that the capital buffer just got much larger. The subsidy can continue longer. The question of whether it can end profitably remains unanswered.

The Copilot pricing data point + GPT-5.5 pricing = the token economics picture

Microsoft’s $30/$70 pooled credits per Copilot seat + GPT-5.5 Pro pricing at $30/$180 per 1M tokens + Anthropic’s $5/$25 Opus 4.7 pricing now give us enough data points to map the enterprise agent cost surface. At GPT-5.5 Pro rates, a Business Copilot seat’s $30 monthly credit buys roughly 167K output tokens — about 2-3 substantive agent sessions. At standard GPT-5.5 rates, it buys 1M output tokens — more usable, but still constrained.

The implication for anyone building open-source coding agents: the model cost is the primary constraint on agent utility, and it varies 6x between standard and reasoning tiers from the same vendor. Local models (Qwen3.6-27B, Gemma 4 31B) eliminate this cost entirely for coding tasks where they’re competitive, at the cost of lower benchmark performance. The cost-quality tradeoff is now quantifiable.

Security

No new CVEs or advisories for tracked dependencies. The five-dimension Claude Code security surface remains unchanged — unpatched credential exfiltration chain (CVE-2026-35020/35021/35022) is still the primary open issue.

What to watch

1. GPT-5.5 adoption in Codex — does the terminal-bench advantage translate to better agent performance in practice?
2. Anthropic response — Google just invested $40B in them and OpenAI just shipped a competitive model. Does Claude Code get a model bump or feature sprint?
3. Qwen3.6-27B local evaluation — dense model vs MoE model on practical coding tasks with Apple Silicon hardware constraints
4. huihui-ai Huihui4 — original model or abliteration? The naming shift matters.
5. aube ecosystem adoption — mise now defaults to aube as npm backend. First real-world migration stories.
6. Post-deadline developer sentiment — Copilot data training activated yesterday, billing announced two days ago. Any organized response?