DeepSeek V4 Rebuffs US Accusations.

Mazen Al Sabbagh

— 10 min read
Share

TL;DR

  • DeepSeek V4 deployed: China's DeepSeek released V4-Pro and V4-Flash, open-source models optimized for Huawei chips, less than 24 hours after the White House accused Beijing of "industrial-scale" AI theft. Benchmarks position V4-Pro as a leader in math, though it lags on general world knowledge. (Reuters, Bloomberg, OfficeChai)
  • GPT-5.5 "Spud" reception: OpenAI’s latest flagship is registering "measured enthusiasm" post-launch. Artificial Analysis places it at 60 on the Intelligence Index, securing the lead, yet API access is pending, and Claude Opus 4.7 retains its edge on SWE-Bench Pro. (Artificial Analysis)
  • Cohere acquires Aleph Alpha: Valuing the combined entity at ~$20 billion, Cohere's acquisition of Germany's Aleph Alpha — bolstered by a $600 million investment from Schwarz Group — establishes a transatlantic "sovereign AI" powerhouse. The strategic intent is clear: to court global customers seeking alternatives to Silicon Valley's AI hegemony. (NYT, Reuters)
  • Anthropic-White House détente: The Trump administration has reversed course on its dispute with Anthropic. The DOJ has moved to pause its appeal, the White House has engaged CEO Dario Amodei, and President Trump has publicly praised Anthropic leadership. This strategic pivot follows earlier federal agency demand for Anthropic models. (Politico)
  • LATE BREAK: Google commits $40B to Anthropic: Google is investing up to $40 billion in Anthropic — an immediate $10 billion cash injection at a $350 billion valuation, with an additional $30 billion tied to performance, alongside 5 GW of Google Cloud compute over five years. This follows Amazon's $25 billion commitment, positioning Anthropic with ~$65 billion in strategic capital this week. (Bloomberg, CNBC)

Lead Story: DeepSeek V4 Lands the Morning After Washington's AI Theft Accusation

The timing represents either pure coincidence or a deliberate strategic declaration. On Thursday, the White House Office of Science and Technology Policy (OSTP) issued memo NSTM-4, directly accusing "foreign entities, principally based in China," of systemic, "industrial-scale" AI theft via distillation from models like Claude and GPT. DeepSeek's response arrived not in communiqué, but in code: preview versions of DeepSeek-V4-Pro and DeepSeek-V4-Flash, its long-anticipated flagship successors to V3.2, dropped Friday morning.

Both models are open-source, tailored for Huawei chips, unequivocally signaling China's escalating semiconductor autonomy amidst tightening US export controls. V4-Pro integrates 1.6 trillion total parameters with 49 billion activated per token and features a native one-million-token context window. The Flash variant emphasizes cost efficiency, operating at 10% of V3.2's FLOPs in long-context workloads, per DeepSeek's technical documentation. The models leverage a Hybrid Attention Architecture for enhanced long-context retention.

Per Simon Willison, who tested the models within hours of release, both benchmark robustly, and their "prices put them both as the cheapest in their respective categories by a solid margin." Midday benchmark details from OfficeChai and TechCrunch are now clarifying the picture: V4-Pro leads open models in math benchmarks (IMOAnswerBench 89.8), surpassing Claude's 75.3 and Gemini's 81.0, though trailing GPT-5.4 at 91.4. DeepSeek concedes V4-Pro lags Gemini-3.1-Pro on broader "rich world knowledge." Fortune characterized the release as "rock-bottom prices and close integration with Huawei's chips." Late Friday, VentureBeat underscored this, noting V4-Pro delivers "near state-of-the-art intelligence at 1/6th the cost of Opus 4.7 [and] GPT-5.5." Decrypt highlighted V4-Pro's 98% cost advantage over GPT-5.5 Pro.

Meanwhile, Tencent and Alibaba are in advanced discussions to invest in DeepSeek, with valuations now reportedly exceeding $20 billion, up from the $10 billion floor seen days prior. Tencent's proposed 20% stake faces reluctance from DeepSeek, per Bloomberg. The V4 launch arrives as DeepSeek's initial external funding round is actively oversubscribed.

The White House memo explicitly detailed that distillation campaigns "strip security protocols" and dismantle "ideologically neutral" mechanisms from copied models. China's foreign ministry dismissed the accusation as "baseless." The geopolitical implications of today's V4 deployment are undeniable.

In Other News

GPT-5.5 Day Two: Benchmarks Hold, API Still Pending. A day after OpenAI's launch of GPT-5.5 — its inaugural fully retrained base model since GPT-4.5 — initial benchmark figures are solidifying under independent review. Artificial Analysis confirmed GPT-5.5 tops its Intelligence Index at 60, a three-point lead over Claude Opus 4.7 and Gemini 3.1 Pro at 57. Terminal-Bench 2.0 remains its strongest card at 82.7%, a 13-point lead over Opus 4.7. However, OpenAI declined to publish SWE-Bench Pro numbers — where Claude Opus 4.7 leads at 64.3% — and API access for cybersecurity applications remains delayed. Developer sentiment is "measured, not euphoric," appreciating token efficiency (~40% fewer tokens per Codex task), but reserving judgment pending independent LiveBench and HumanEval+ evaluations. Priced at $5/$30 per million tokens—double GPT-5.4—it is clearly positioned as an enterprise agentic model, not a budget offering.

Cohere Acquires Aleph Alpha in $20 Billion Sovereign AI Play. The most significant AI M&A event of the year unfolded Friday morning: Canada's Cohere is acquiring Germany's Aleph Alpha, creating a transatlantic AI entity valued at approximately $20 billion, per Sifted and CNBC. Schwarz Group, the retail conglomerate and a primary Aleph Alpha backer, is injecting $600 million into Cohere’s forthcoming Series E. The combined entity will operate under the Cohere name with CEO Aidan Gomez at the helm, headquartered across Toronto and Heidelberg. This transaction is explicitly framed as "sovereign AI for the world"—per the companies' own announcement—targeting governments, regulated industries, and enterprises seeking AI infrastructure independent of the Anthropic-OpenAI-Google axis. The NYT characterized it as a move "aimed at customers uneasy about the dominance of American companies in artificial intelligence." In a week witnessing Chinese AI self-sufficiency from DeepSeek and White House accusations of distillation theft, a $20 billion bet on non-US, non-China AI sovereignty is remarkably timely.

LATE BREAK: Google Commits Up to $40 Billion to Anthropic. The largest AI investment to date materialized late Friday: Google is committing up to $40 billion to Anthropic, per Bloomberg and CNBC. The investment comprises an immediate $10 billion cash injection at a $350 billion valuation—consistent with Anthropic's February round—with up to $30 billion more contingent on defined performance milestones. Additionally, Google Cloud will provision five gigawatts of computing capacity over a five-year period. The NYT noted the deal "tightens the companies' relationship even as they compete in some areas." Bloomberg reported that Google executives were "anxious about being outpaced in AI coding," a rapidly expanding segment where Anthropic’s Claude Code maintains a commanding lead. The timing is striking: this follows Amazon's commitment of up to $25 billion to Anthropic just three days prior, meaning the Claude maker has secured approximately $65 billion in strategic investment commitments in a single week. TechCrunch framed the deal as driven by "surging demand for Claude Code" and the imperative to scale infrastructure for Anthropic's Cowork agent. Google previously held a ~14% stake from prior ~$3 billion investments. On secondary markets, Anthropic's valuation has reportedly exceeded $1 trillion—nearly triple the primary round price.

The Trump Administration's Anthropic U-Turn Is Now Complete. What began as an administration-led blacklisting has concluded as a strategic courtship. Per Politico, the DOJ filed to freeze its own appeal of the March injunction; the White House held a cordial meeting with CEO Dario Amodei; and President Trump stated on CNBC that Anthropic executives are "very smart people." The catalyst is pragmatic: federal agencies—including the NSA, which leveraged Mythos via an ODNI loophole despite a Pentagon blacklist—require the model for cyber defense applications. A negotiated resolution appears to be materializing.

Meta's Layoffs Arrive With a Surveillance Footnote. Meta will execute a 10% workforce reduction—approximately 8,000 employees—effective May 20, concurrently eliminating 6,000 open roles. This move offsets $135 billion in 2026 AI capital expenditure. The announcement coincided with Meta's disclosure of logging employee keystrokes, mouse movements, and screenshots to train its AI models. This confluence of mass layoffs and AI-powered employee surveillance generated particularly acute headlines. Meta reports Q1 earnings April 29.

X / Social Pulse

Within hours of the V4 preview release, Simon Willison published a comprehensive analysis, characterizing both models as competitive, "almost on the frontier, a fraction of the price," and noting their aggressively undercut pricing. On Bluesky and X, his canonical 'pelican' image-generation test yielded competent, yet visibly 'angry-looking' pelicans from both V4-Flash and V4-Pro. Community member @jukan05 captured the sentiment: "What kind of magic did DeepSeek pull off this time? With V4, they seem to be back at SOTA again." The significant 484-day gap between V3 and V4, highlighted by a DeepSeek engineer, frames the release as a meticulously crafted effort.

Anthropic's post-mortem on Claude Code's degradation, attributing it to a reasoning effort downgrade, a verbosity prompt change, and a caching defect, received a mixed reception. The Register headlined it as Anthropic admitting it "dumbed down Claude with upgrades," while another piece critiqued Opus 4.7 as an "overzealous query cop" due to escalating refusal rates from its Acceptable Use Classifier. DevOps.com characterized the issue as a "code quality conundrum" persisting despite Anthropic's pledges for stricter controls. Developer sentiment on X remains polarized: some accept the explanation, others perceive a recurrent pattern of shipping rapidly and debugging in production.

A widely circulated analytical post by ThePlanetTools argues that "independent benchmarks you see published this week are measuring models released last month." It highlights that the rapid, six-week release cadence from frontier labs structurally lags third-party evaluations by two to three cycles. The recommendation for enterprise buyers is direct: "pilot GPT-5.5 Business alongside Claude Opus 4.7 and Gemini 3.1 Pro, and let your engineering team vote with their terminals." This perspective—emphasizing distribution and workflow integration over incremental benchmark deltas—is gaining considerable traction within developer communities as the top three labs demonstrate increasing capability convergence.

One to Watch

FISA Section 702 Expires April 30 — and It's Not Just a Surveillance Story. Congress provisionally extended FISA Section 702 until April 30 via a 10-day stopgap, after failing to pass either a five-year or eighteen-month renewal on April 17. The impasse stems from familiar points: privacy advocates demand warrant requirements, certain factions of libertarian Republicans seek to attach the SAVE America Act, while the administration cites battlefield imperatives for a clean extension. Given the NSA's active deployment of Anthropic's Mythos for cyber defense under an ODNI loophole, the entanglement between FISA's legal architecture and government frontier AI deployments is becoming structurally critical. A lapse, or a hastily negotiated resolution, holds significant downstream implications for the governance of classified AI deployments. Six days remain critical.

Quick Hits

  • Anthropic has expanded Claude's connected apps to over 15 consumer services including Spotify, Uber, Instacart, TurboTax, AllTrails, and Audible, transitioning Claude from solely workplace integrations into broader consumer-facing applications. (The Verge, PCMag)
  • Global equity fund inflows reached a 17-month peak in the week through April 22, propelled by AI optimism and robust Q1 bank earnings, per LSEG Lipper data. AI-linked stocks now constitute 45% of S&P 500 market capitalization, a significant escalation from 25% at ChatGPT's debut. (Reuters, Economic Times)
  • Intel’s Q1 earnings exceeded expectations decisively: reporting $13.6B revenue (+7% YoY) and Data Center and AI growth of 22%, driving a +16% after-hours stock surge; CEO Lip-Bu Tan confirmed the Terafab partnership with SpaceX/xAI/Tesla during the earnings call. (TNW)
  • Jury selection for Musk v. Altman commences Monday, April 27, in Oakland — Musk is seeking damages exceeding $150 billion and Altman's removal from OpenAI's board; key witnesses include Satya Nadella and Mira Murati; Judge Gonzalez Rogers has reserved four weeks for the trial. (NYT)
  • OpenAI's consumer video platform, Sora, ceases operations Sunday, April 26 — the application closes two days from now after an estimated $1 million daily compute expenditure; its API will remain active until September 24. (Wikipedia))

The week concluding April 24 will be recognized as the point where every significant AI narrative converged. We witnessed simultaneous frontier model launches from the West (GPT-5.5) and East (DeepSeek V4), the latter a pointed response to a White House accusation of industrial-scale AI theft. A $20 billion transatlantic merger established a "sovereign AI" counter-narrative (Cohere-Aleph Alpha), and then, in an unprecedented capital allocation, Google committed up to $40 billion to Anthropic, augmenting Amazon's $25 billion pledge just 72 hours prior. Anthropic now commands an extraordinary market position: a trillion-dollar secondary valuation, $65 billion in new strategic capital from the dominant cloud providers, a White House that shifted from blacklisting to active engagement, and a model (Mythos) drawing attention from the NSA, central banks, and the cyber underground. The competitive intensity among frontier labs is acute, with Artificial Analysis separating the top three by merely three Intelligence Index points, while DeepSeek demonstrates comparable capability at a significantly lower cost. As earnings season commences next week with Alphabet, Meta, and Microsoft reporting, the industry's balance sheets must now rationalize these aggressive technological bets.

Sources

Lock in. M. mazen@thorterminal.com

Read more