Lock-in day: OpenAI consolidates, DeepSeek picks Huawei, Google backs Anthropic

TL;DR

OpenAI sunsets every Codex SKU into GPT-5.5 by July 2026, doubling API prices while trailing Claude on long migrations.
DeepSeek V4 ships open-weight on Huawei Ascend silicon and immediately draws US accusations of distilling 16M Claude exchanges.
Google commits up to $40B to Anthropic days after Amazon’s round, making the lab uniquely anchored to both AWS and GCP.
Samsung executives reportedly fear the handset division’s first-ever annual loss as AI memory demand inflates its own HBM and DRAM costs.
South Korean prosecutors seek five years for a man who fabricated AI-generated sightings of an escaped zoo wolf.

The headlines today aren’t about benchmarks or demos — they’re about contracts and commitments. OpenAI is collapsing the Codex line into GPT-5.5 and shutting every codex SKU by July, betting its agentic-coding future on a single flagship even as Claude still leads on the workloads that matter most to enterprises. DeepSeek shipped V4 not on NVIDIA but on Huawei’s Ascend stack, a hardware choice that became a State Department story within twenty-four hours. And Google is putting up to $40 billion into Anthropic days after Amazon’s own round, leaving the lab with a dual-cloud patron structure no other frontier company has.

Each move tightens a constraint that will be expensive to undo: a product lineup, a silicon roadmap, a capital partner. The round-ups underline the same theme from the periphery — Anthropic deepening enterprise channels in Japan, Samsung’s handset margins squeezed by the memory boom its rivals are driving. Capability races get the airtime; today, the structural bets are the story.

Google doubles down on Anthropic with up to $40B, days after Amazon

Source: ars-technica-ai · published 2026-04-24

TL;DR

Google will put up to $40 billion into Anthropic, following a fresh Amazon round closed days earlier.
Anthropic now has both major US cloud providers as anchor investors and compute suppliers — a uniquely dual-patron structure.
The check size locks in TPU and Trainium demand for years and pushes Anthropic into OpenAI-Microsoft territory for capital intensity.

The headline number

Google has committed as much as $40 billion to Anthropic, Ars Technica reports, arriving on the heels of a separate Amazon investment closed earlier in the same week. Anthropic is now the only frontier lab with two hyperscalers writing checks of this magnitude into the same cap table at roughly the same time.

The “as much as” framing matters. It typically signals a tranched or milestone-gated commitment rather than a single wire transfer, and the exact split between cash and committed cloud credits hasn’t been independently reported. But even at the low end of plausible drawdown schedules, this is one of the largest single private investments ever made in any company, in any sector.

Why both clouds, and why now

The structurally interesting part isn’t the number, it’s the shape. OpenAI is fused to Microsoft Azure; xAI runs on its own Colossus build-out; Meta and Google train in-house. Anthropic alone has played both sides — Claude runs on AWS Trainium and Google TPUs — and now both clouds have written checks large enough to make that dual-sourcing structurally permanent.

flowchart LR
    G[Google] -- up to $40B --> A((Anthropic))
    AM[Amazon] -- new round --> A
    A -- training/inference --> TPU[Google TPUs]
    A -- training/inference --> TR[AWS Trainium]
    A -- Claude API --> CUST[Enterprise customers]
    G -. cloud revenue .- TPU
    AM -. cloud revenue .- TR

That matters for three reasons:

Compute hedging. Neither AWS nor Google can afford to be cut off from the most-cited non-OpenAI frontier model without handing the other a decisive enterprise-AI advantage.
Chip validation. A commitment of this size is the strongest available endorsement that TPUs and Trainium are credible alternatives to Nvidia at training scale — something both Google and Amazon badly need to demonstrate to other customers.
Revenue round-tripping. A large fraction of these dollars will return to the investor as cloud spend. The accounting question of how much of Anthropic’s “revenue” is hyperscaler-funded compute credits is now unavoidable.

What we don’t yet know

Beyond the headline figure, the post-money valuation, governance rights, exclusivity carve-outs against the existing Amazon relationship, and the cash-versus-credit split haven’t surfaced in public reporting. The “up to” qualifier in particular leaves wide latitude — a $40B ceiling could resolve to substantially less actual capital deployed depending on milestones we can’t see.

What is clear: Anthropic, founded in 2021 by ex-OpenAI researchers, has gone from a safety-focused research spinout to a company that two of the world’s three largest cloud providers have decided they cannot afford to lose.

The takeaway

The frontier-lab landscape has hardened into a duopoly of capital structures: OpenAI inside Microsoft, and Anthropic balanced between Google and Amazon. Everyone else — Mistral, Cohere, AI21, the open-weights ecosystem — is now competing for a different and much smaller pool of capital. If you were waiting for a sign that the “many frontier labs” thesis is dead, this is it. The interesting AI competition over the next 24 months isn’t model-versus-model; it’s whether Anthropic’s two-cloud structure proves to be leverage or a coordination tax.

OpenAI folds Codex into GPT-5.5 — and inherits its problems

Source: simon-willison · published 2026-04-25

TL;DR

OpenAI killed the standalone Codex model line; GPT-5.5 absorbs agentic coding and computer use as core capabilities.
Every gpt-5-codex and gpt-5.1-codex SKU shuts down July 23, 2026, at double the previous API price.
GPT-5.5 wins Terminal-Bench (82.7%) but trails Claude Opus 4.7 on SWE-bench Pro and produces 2x more cleanup work on long migrations.
The “Codex” brand survives as an app — Chronicle screen memory and AGENTS.md ship with a thin safety story.

What “unification” actually buys OpenAI

Romain Huet’s tweet (“no separate coding line anymore”) sounds like SKU cleanup. It isn’t. OpenAI is sunsetting gpt-5-codex and the gpt-5.1-codex family entirely on July 23, 2026 ¹, which forces every developer on the dedicated coding API onto GPT-5.5’s general endpoint at roughly $5/$30 per million tokens — double GPT-5.4. OpenAI’s defense is a claimed ~40% token-efficiency gain ²; the practical effect is that the cheapest way to do agentic coding from OpenAI is now the same model you’d use for anything else, billed accordingly.

The “Codex” name doesn’t disappear — it migrates up the stack to the CLI, VS Code extension, and macOS superapp covered in the Latent Space recap. Codex is now an application surface, not a model.

The benchmark story is genuinely split

Benchmark	GPT-5.5	Claude Opus 4.7
Terminal-Bench 2.0	82.7%	—
OSWorld-Verified	78.7%	—
SWE-bench Pro	58.6%	64.3%
400-file migration cleanup pass	13.1%	5.8%

GPT-5.5 dominates the surfaces OpenAI optimized for — terminal use and computer-use agents ³. But MindStudio’s long-horizon evaluation found GPT-5.5 finishes 400+ file migrations 32% faster while producing more than twice Claude’s rate of follow-up fixes, which reviewers blame on “context drift” — GPT reasoning file-locally while Claude re-anchors to the original task ⁴. Faster, but messier on the exact long-running refactors Codex users bought it for.

Chronicle expands the attack surface

The Codex superapp ships Chronicle, a screen-watching memory layer Sam Altman compared to “telepathy.” It stores memories as unencrypted markdown in ~/.codex/memories/ and ships screen captures to OpenAI for OCR ⁵. That’s a fresh indirect-prompt-injection vector that doesn’t require the user to paste anything — merely viewing a hostile page can plant instructions the agent later acts on.

flowchart LR
    A[Untrusted webpage<br/>user merely views] --> B[Chronicle screen capture]
    B --> C[OpenAI OCR]
    C --> D[~/.codex/memories/<br/>unencrypted markdown]
    D --> E{Codex agent}
    F[AGENTS.md<br/>Linux Foundation spec] --> E
    G[Poisoned dependency] -. rewrites .-> F
    E --> H[Shell, browser, files]

The companion primitive, AGENTS.md — now stewarded by the Linux Foundation — has a parallel problem: NVIDIA researchers flagged it as a build-time supply-chain target. OpenAI’s own system card rates GPT-5.5 “High” on cybersecurity risk.

The competitive shockwave

Unification also makes OpenAI a direct competitor to its biggest API customers. Cursor (now ~$2.7B ARR) responded by training Composer 2, claimed at ~72x cheaper than GPT-5.5 Pro though well behind on agentic reasoning ⁶. Google’s $2.4B Windsurf licensing deal and Cognition’s ~$250M acquisition of the IDE shell were the prior chapter of the same realignment: if the frontier model is the coding agent, every IDE either trains its own model or becomes a thin client.

Takeaway

The headline (“no more Codex line”) is tidy. The reality is a pricier model that wins the benchmarks OpenAI chose, regresses on the long-horizon work Codex was bought for, and ships a memory layer whose threat model isn’t yet credible. The unification is real; the safety story is provisional.

DeepSeek V4 lands on Huawei silicon — and on a State Department cable

Source: mit-tech-review-ai · published 2026-04-24

TL;DR

DeepSeek shipped V4-Pro (1.6T-A49B) and V4-Flash (284B-A13B) open-weight, co-optimized for Huawei’s Ascend stack rather than NVIDIA.
A new hybrid attention design claims 1M-token context at ~10% of V3.2’s KV cache and 27% of its inference FLOPs.
Within 24 hours, US officials accused DeepSeek of distilling ~16M Claude exchanges via 24,000 fraudulent accounts.
NIST flags V4 as materially more vulnerable to agent hijacking than US frontier peers.

A frontier model that doesn’t need NVIDIA

The load-bearing fact about DeepSeek V4 isn’t the parameter count — it’s the silicon. V4-Pro is a 1.6-trillion-parameter MoE (49B active) trained and served on Huawei’s CloudMatrix 384 / Ascend 950 stack, making it the first frontier-class open-weight release that routes around the NVIDIA supply chain entirely ⁷. The companion V4-Flash (284B-A13B) is small enough that FP8 distillations are plausibly local-runnable, which is what has practitioners paying attention.

The architecture story is a hybrid attention stack — Compressed Sparse Attention plus Heavily Compressed Attention — that lets V4-Pro hold a default 1M-token window at roughly 10% of V3.2’s KV cache and 27% of its inference FLOPs ⁸. That is a genuine efficiency jump, and it is what makes serving 1.6T parameters on Ascend hardware economically defensible at all.

The 1M context has an asterisk

Independent teardowns are less generous about what “1M tokens” actually buys you. RULER and MRCR-8-needle recall degrades meaningfully past 128k–200k tokens, and one analysis describes the extreme-compression tier as a “blurry statistical soup” that keeps global gist but loses granular detail ⁸. Reasoning chains drift past ~30 tool calls. CFR’s read is similarly sober: DeepSeek still trails US frontier labs by 3–6 months on raw reasoning, and the company itself has acknowledged it cannot serve V4-Pro at scale because of domestic compute shortages ⁹.

The HN reception mirrors that split. Developers call V4 an “insane value deal,” but the dominant complaint is that the official API “fails half the time,” with one widely-upvoted comment arguing leaderboards need a reliability score ¹⁰. Some testers reported V4-Flash beating V4-Pro on agentic benchmarks — almost certainly a provider-stability artifact, not a model fact.

The distillation cable changes the frame

Within 24 hours of release, the White House OSTP and a State Department cable accused DeepSeek of industrial-scale distillation from US frontier models. The specific number, sourced to Anthropic, is concrete enough to matter:

~24,000 fraudulent accounts generating over 16 million exchanges with Claude, targeted at reasoning and coding traces ¹¹.

This reframes the entire launch. The DeepSeek narrative — efficient training, clever architecture, sovereign hardware — now competes with a parallel narrative of IP exfiltration at API scale. The two aren’t mutually exclusive (you can distill teacher traces and engineer a good sparse-attention stack), but the cable lands the same week as the weights, and policy readers will fuse them.

The safety gap nobody is selling

The least-discussed finding may be the most consequential for anyone considering V4 in an agentic deployment. NIST’s CAISI evaluation found DeepSeek models significantly more vulnerable to agent hijacking and malicious instruction injection than US-based frontier peers ¹². Pair that with the 30-tool-call reasoning drift, and “open frontier model on Chinese silicon” stops looking like a drop-in substitute for Claude or GPT in any pipeline that touches untrusted input. The headline is real; the operational caveats are larger than the launch posts admit.

Round-ups

Anthropic and NEC collaborate to build Japan’s largest AI engineering workforce

Source: anthropic-news

Anthropic is partnering with NEC to train what the two companies pitch as Japan’s largest AI engineering workforce, deepening Claude’s enterprise foothold in the country through NEC’s systems integration channels and internal developer upskilling programs.

Report: Samsung execs worried company could lose money on smartphones for the first time

Source: ars-technica-ai

Samsung executives reportedly fear the smartphone division could post its first-ever annual loss, as the AI-driven memory boom drives up HBM and DRAM prices that Samsung’s own handset business has to pay, squeezing margins on Galaxy devices.

ChatGPT’s Nano Banana

Source: bens-bites

Ben’s Bites benchmarks ChatGPT’s new image model — informally dubbed its ‘Nano Banana’ — against popular design tools, testing how it stacks up on common creative workflows like mockups, edits, and brand-consistent generations.

An update on our election safeguards

Source: anthropic-news

Anthropic published an update on the election integrity safeguards it has applied to Claude, covering policy enforcement, misuse monitoring, and partnerships meant to prevent the model from being weaponized for voter manipulation or fabricated political content.

The people do not yearn for automation

Source: simon-willison

Simon Willison highlights Nilay Patel’s Verge essay arguing that ChatGPT usage keeps climbing while public sentiment toward AI sours because ‘software brain’ executives reduce human experience to automatable information flows — the same instinct that has kept smart-home adoption niche for over a decade.

AIE Europe Debrief + Agent Labs Thesis: Unsupervised Learning x Latent Space Crossover Special (2026)

Source: latent-space

Latent Space and Unsupervised Learning’s crossover episode debriefs the AI Engineer Europe conference and lays out an ‘Agent Labs’ thesis on the emerging crop of agent-focused startups; recorded before the Cursor–xAI deal landed.

Man faces 5 years in prison for using AI to fake sighting of runaway wolf

Source: ars-technica-ai

South Korean prosecutors are seeking up to five years in prison for a man who fabricated AI-generated photos of Sejong, an escaped zoo wolf that captivated the country, after sharing the fake sightings online for amusement during the manhunt.

i10x migration guide — https://i10x.ai/news/openai-codex-api-shutdown-migrate-to-gpt-5-5

intermediate models—including gpt-5-codex and various gpt-5.1-codex variants—are scheduled for a complete shutdown on July 23, 2026

↩
Artificial Intelligence News — https://www.artificialintelligence-news.com/news/gpt-5-5-is-openais-most-capable-agentic-ai-model-yet-at-twice-the-api-price/

GPT-5.5 is OpenAI’s most capable agentic AI model yet at twice the API price

↩
MindStudio benchmark roundup — https://www.mindstudio.ai/blog/gpt-5-5-vs-claude-opus-4-7-agentic-coding

GPT-5.5 reached 82.7% on Terminal-Bench 2.0 … but Claude Opus 4.7 still leads SWE-bench Pro at 64.3% vs 58.6%

↩
MindStudio long-horizon eval — https://www.mindstudio.ai/blog/claude-vs-gpt-agentic-coding-comparison

GPT-5.5 required a 13.1% cleanup pass compared to just 5.8% for the latest Claude models … attributed to context drift

↩
BigGo / Latent Space recap on Chronicle — https://finance.biggo.com/s/Copilot%20Task

memories are stored as unencrypted markdown files … Chronicle increases susceptibility to indirect prompt injection

↩
Capwave on Cursor’s response — https://capwave.ai/blog/windsurf-cognition-and-google-what-just-happened-in-ai

Cursor has begun training its own frontier-level model, Composer 2 … roughly 72x cheaper than GPT-5.5 Pro

↩
Tom’s Hardware — https://www.tomshardware.com/tech-industry/artificial-intelligence/deepseek-launches-1-6-trillion-parameter-v4-on-huawei-chips-as-us-escalates-ai-theft-accusations

DeepSeek launches 1.6-trillion-parameter V4 on Huawei chips as US escalates AI theft accusations

↩
Medium — ‘The 1M Context Lie’ (Aditya J) — https://medium.com/@adityaj5400/the-1m-context-lie-why-v4s-hybrid-attention-is-the-death-of-the-8-h100-standard-d2e4066960d4

DeepSeek-V4-Pro to operate at 1M tokens while requiring only 10% of the KV cache and 27% of the inference FLOPs used by its predecessor, V3.2

↩ ↩²
Council on Foreign Relations — https://www.cfr.org/articles/deepseek-v4-signals-a-new-phase-in-the-u-s-china-ai-rivalry

DeepSeek leads the open-weight field but still trails the leading US frontier models by roughly 3 to 6 months in raw reasoning

↩
Hacker News thread (item 47894988) — https://news.ycombinator.com/item?id=47894988

leaderboards should include a ‘reliability score’ because frequent API errors often mask the model’s actual reasoning quality

↩
StartupFortune (on State Dept cable) — https://startupfortune.com/the-us-just-sent-a-global-diplomatic-warning-about-chinese-ai-theft-and-the-case-against-deepseek-is-more-specific-than-the-headlines-suggest/

Anthropic reported that DeepSeek and other firms utilized approximately 24,000 fraudulent accounts to generate over 16 million exchanges with the Claude model

↩
NIST/CAISI evaluation — https://www.nist.gov/news-events/news/2025/09/caisi-evaluation-deepseek-ai-models-finds-shortcomings-and-risks

DeepSeek models are significantly more vulnerable to ‘agent hijacking’ and malicious instructions compared to U.S.-based frontier models

↩

Lock-in day: OpenAI consolidates, DeepSeek picks Huawei, Google backs Anthropic

Lock-in day: OpenAI consolidates, DeepSeek picks Huawei, Google backs Anthropic

TL;DR

Google doubles down on Anthropic with up to $40B, days after Amazon

TL;DR

The headline number

Why both clouds, and why now

What we don’t yet know

The takeaway

OpenAI folds Codex into GPT-5.5 — and inherits its problems

TL;DR

What “unification” actually buys OpenAI

The benchmark story is genuinely split

Chronicle expands the attack surface

The competitive shockwave

Takeaway

Further reading

DeepSeek V4 lands on Huawei silicon — and on a State Department cable

TL;DR

A frontier model that doesn’t need NVIDIA

The 1M context has an asterisk

The distillation cable changes the frame

The safety gap nobody is selling

Further reading

Round-ups

Anthropic and NEC collaborate to build Japan’s largest AI engineering workforce

Report: Samsung execs worried company could lose money on smartphones for the first time

ChatGPT’s Nano Banana

An update on our election safeguards

The people do not yearn for automation

AIE Europe Debrief + Agent Labs Thesis: Unsupervised Learning x Latent Space Crossover Special (2026)

Man faces 5 years in prison for using AI to fake sighting of runaway wolf

Jack Sun, writing.

Lock-in day: OpenAI consolidates, DeepSeek picks Huawei, Google backs Anthropic

TL;DR

Google doubles down on Anthropic with up to $40B, days after Amazon

TL;DR

The headline number

Why both clouds, and why now

What we don’t yet know

The takeaway

OpenAI folds Codex into GPT-5.5 — and inherits its problems

TL;DR

What “unification” actually buys OpenAI

The benchmark story is genuinely split

Chronicle expands the attack surface

The competitive shockwave

Takeaway

Further reading

DeepSeek V4 lands on Huawei silicon — and on a State Department cable

TL;DR

A frontier model that doesn’t need NVIDIA

The 1M context has an asterisk

The distillation cable changes the frame

The safety gap nobody is selling

Further reading

Round-ups

Anthropic and NEC collaborate to build Japan’s largest AI engineering workforce

Report: Samsung execs worried company could lose money on smartphones for the first time

ChatGPT’s Nano Banana

An update on our election safeguards

The people do not yearn for automation

AIE Europe Debrief + Agent Labs Thesis: Unsupervised Learning x Latent Space Crossover Special (2026)

Man faces 5 years in prison for using AI to fake sighting of runaway wolf

Footnotes

Jack Sun, writing.