Search

DeepSeek-V4 and GPT-5.5 invade IDEs, Cohere merges with Aleph Alpha, Claude Code v2.1.119

DeepSeek-V4 and GPT-5.5 invade IDEs, Cohere merges with Aleph Alpha, Claude Code v2.1.119

April 25, 2026 is packed with major announcements on frontier models and developer tools. DeepSeek-V4 is launched open-source and available free on NVIDIA Blackwell. GPT-5.5 enters general availability in GitHub Copilot and the OpenAI API. The Cohere + Aleph Alpha merger, backed by USD 600 million from Schwarz Group, lays the groundwork for transatlantic sovereign AI. On the tools side, Claude Code v2.1.119 and Codex’s Auto-review mode extend the autonomy of development agents.


DeepSeek-V4 available everywhere

Launch and open-source

24 April — DeepSeek simultaneously announces DeepSeek-V4-Pro and DeepSeek-V4-Flash. V4-Pro is a 1.6 trillion-parameter mixture-of-experts model (49 billion active), with a one-million-token context window and performance claimed to be comparable to the best closed models. V4-Flash, more compact (284B/13B active), targets low-latency use cases. Both models are open-source from day one, with API and demo available immediately, and the technical report published on Hugging Face.

🔗 DeepSeek-V4 announcement

API promotion and integrations

25 April — DeepSeek announces a -75% promotion on the V4-Pro API until May 5, 2026 (3:59 PM UTC). The Claude Code, OpenCode, and OpenClaw integrations have been updated to support the new model.

🔗 DeepSeek-V4-Pro API promotion

DeepSeek-V4-Pro on NVIDIA Blackwell, free of charge

24 April — NVIDIA makes DeepSeek-V4-Pro available for free via the NVIDIA NIM API interface on Blackwell, on build.nvidia.com. The announcement generated 160,000 views. NVIDIA also publishes the first performance curves (Pareto frontier) for DeepSeek-V4-Pro on Blackwell Ultra with vLLM — an early reference point for high-performance deployments.

25 April — NVIDIA also highlights the limits of classic inference for development agents: “Traditional inference wasn’t built for agentic coding”, referring to the hundreds of API calls generated by modern agentic tools.

🔗 DeepSeek-V4-Pro on NVIDIA NIM 🔗 Blackwell Ultra Day 0


GPT-5.5 leaves preview

General availability in GitHub Copilot

24 April — GPT-5.5 is being rolled out gradually in GitHub Copilot for Pro+, Business, and Enterprise plans. Availability covers VS Code, Visual Studio, the Copilot CLI, the cloud agent, github.com, the mobile app (iOS and Android), JetBrains IDEs, Xcode, and Eclipse. The promotional multiplier is set at 7.5×. Enterprise and Business admins must enable the policy in settings to benefit from it.

🔗 GitHub Copilot changelog — GPT-5.5 GA

Developer API access

24 April — The day after the public launch, OpenAI opens access to GPT-5.5 in the API. The model is available via the Responses API and the Chat Completions API, with a one-million-token context window. The GPT-5.5-Pro variant, for high-precision work, is available only via the Responses API.

“GPT-5.5 is available in the Responses and Chat Completions APIs with a 1M context window. GPT-5.5-pro is also available in the Responses API for higher-accuracy work.”

🔗 OpenAI API announcement

GPT-5.5 on Perplexity Max and Personal Computer

24 April — GPT-5.5 is available to Max subscribers on Perplexity and deployed as the default orchestration model in Personal Computer for Pro and Max plans.

🔗 Perplexity announcement

Bio Bug Bounty — biosafety security program

23 April — OpenAI opens a bug bounty program dedicated to GPT-5.5 biosafety. AI security or biosafety researchers are invited to find a universal jailbreak bypassing the model’s biological safeguards. The main reward is USD 25,000 for the first success. Applications are open until June 22, 2026; testing will run from April 28 to July 27, 2026, exclusively on GPT-5.5 in Codex Desktop, by invitation with a confidentiality agreement.

🔗 GPT-5.5 Bio Bug Bounty


Developer tools: extended autonomy

Claude Code v2.1.119

25 April — Anthropic releases Claude Code v2.1.119, a substantial CLI update with more than forty changes.

DomainChange
Configpersistent /config in ~/.claude/settings.json
PR--from-pr supports GitLab, Bitbucket, GitHub Enterprise
HooksPostToolUse + duration_ms field
PowerShellAuto-approval in permission mode
MCPParallel subagent server connection
Fixed bugs40+

The persistence of /config settings is the most visible change: theme preferences, editor mode, or verbose level survive restarts. The --from-pr setting now accepts GitLab merge-request, Bitbucket pull-request, and GitHub Enterprise URLs, extending the workflow to teams that do not use GitHub.com. The PostToolUse and PostToolUseFailure hooks now receive the duration_ms field, useful for CI/CD monitoring. MCP servers are now connected in parallel, reducing startup times for multi-server workflows.

🔗 Claude Code CHANGELOG

Codex Auto-review — extended autonomy with a safety net

24 April — OpenAI announces Auto-review, a new execution mode for Codex. This mode allows Codex to make progress on long tasks without asking for approval at every step. A separate agent evaluates high-risk steps before execution, making it possible to streamline long testing, compilation, and automation workflows without sacrificing security.

🔗 Codex Auto-review

Copilot for JetBrains: Inline Agent Mode in preview

24 April — The Copilot plugin update for JetBrains IDEs brings several features: the inline agent mode in public preview (Shift+Ctrl+I or Shift+Cmd+I shortcut), improved Next Edit Suggestions (NES) with watermark previews and remote changes, and global auto-approval for agent tool calls.

🔗 JetBrains Copilot changelog

NVIDIA Dynamo — inference redesigned for agents

25 April — NVIDIA presents Dynamo, a redesign of the inference stack to match the workload profiles of agentic tools. Agents like Claude Code, Codex, or Copilot chain together hundreds of API calls per session with recomposed context at each step, creating bottlenecks that drive up cost per token. Dynamo combines four components: KV cache-aware routing, agent-oriented scheduling, multi-level caching, and unified orchestration. NVIDIA announces up to 7× more throughput with higher cache hit rates and reduced latency.

🔗 NVIDIA Dynamo — agentic inference


Sovereignty and enterprise partnerships

Cohere + Aleph Alpha: transatlantic merger with USD 600 million from Schwarz Group

24 April — Cohere (Canada) and Aleph Alpha (Germany) announce their merger project. Schwarz Group — the parent group of Lidl and Kaufland — invests USD 600 million (about EUR 500 million) in structured financing for Cohere’s Series E. The combined platform will be hosted on STACKIT, Schwarz Digits’ sovereign cloud.

“Sovereign AI for the world. Cohere & Aleph Alpha form transatlantic AI powerhouse anchored in Canada & Germany!”

L’accord cible les secteurs régulés, les gouvernements et un marché de l’IA souveraine estimé à environ 600 milliards USD. Il reste conditionnel à l’approbation des actionnaires d’Aleph Alpha.

Anthropic and NEC: first global Japanese partnership

24 April — Anthropic announces a strategic partnership with NEC Corporation, which becomes Anthropic’s first global Japanese partner. NEC will deploy Claude to around 30,000 employees of the group worldwide.

AspectDetail
Affected employees~30,000 (global NEC Group)
Deployed productsClaude, Claude Opus 4.7, Claude Code, Claude Cowork
Target sectorsFinance, manufacturing, cybersecurity, local government
ProgramNEC BluStellar Scenario

Internally, NEC is setting up a Center of Excellence to train a large-scale AI engineering team, as part of the “Client Zero” initiative. For its clients, NEC and Anthropic will jointly develop solutions for the finance, manufacturing, and Japanese local government sectors.

🔗 Anthropic and NEC

Meta partners with AWS for agentic AI at billions of users scale

24 April — Meta announces an agreement with AWS to integrate tens of millions of Graviton5 cores into its infrastructure. The goal is to support the CPU-intensive workloads of agentic AI designed for billions of users.

🔗 Meta × AWS Graviton5


Gemini: product and research updates

Gemini Drops April 2026 — Lyria 3 Pro, Gemini Live v3.1

24 April — Google publishes the 10th edition of Gemini Drops. Lyria 3 Pro lets users create music tracks of up to 3 minutes directly in Gemini, available to Plus, Pro, and Ultra subscribers. Gemini Live v3.1 is 20% faster and offers twice as much remembered context. Personal Intelligence expands internationally (excluding the European Economic Area, Switzerland, the United Kingdom, South Korea, Australia, and Nigeria). The branching conversation feature (branching) is rolling out to 20% of users.

🔗 Gemini Drops April 2026

Gemini Embedding 2 in general availability

22 April — Gemini Embedding 2 reaches general availability (GA) in the Gemini API and Vertex AI. The vector representation model targets semantic search, retrieval-augmented generation (RAG), and classification.

🔗 Gemini Embedding 2 GA

Decoupled DiLoCo — multi-region distributed training

23 April — Google DeepMind publishes Decoupled DiLoCo, a distributed training method over low-bandwidth networks. Gemma 12B was trained across 4 US regions with a mix of TPU6e and TPUv5p. The method opens the way to decentralized model training at global scale, without requiring the high-speed interconnects usually needed.

🔗 Decoupled DiLoCo — Google DeepMind


Alternative models: Qwen and Grok

Qwen3.6-27B — flagship dense model for agentic coding

22 April — Alibaba publishes Qwen3.6-27B, a dense 27-billion-parameter open-source model under the Apache 2.0 license. Despite its compact size, it outperforms Qwen3.5-397B-A17B — a 397-billion-parameter MoE model with 17 billion active parameters — on the main agentic coding benchmarks, with a SWE-Bench Verified score of 77.2% versus 76.2% for its predecessor. Three angles are highlighted in the announcement: agentic coding that outperforms the previous generation on all major benchmarks, strong reasoning in text and multimodal, and dense deployment without MoE complexity.

The model supports both thinking and non-thinking modes in the same checkpoint. It is available on Hugging Face (Qwen/Qwen3.6-27B, FP8 variant included) and ModelScope, with a dedicated technical blog and Github. The announcement generated 3.5 million views on X.

🔗 Qwen3.6-27B announcement

Qwen-Image-2.0-Pro — #9 global Text-to-Image

25 April — Alibaba Qwen releases Qwen-Image-2.0-Pro, which reaches 9th place worldwide in the Text-to-Image Arena ranking and 6th place in portrait. The model is available via the Alibaba Cloud API and ModelScope.

🔗 Qwen-Image-2.0-Pro

Grok Voice Think Fast 1.0 — #1 Tau Voice Bench

23 April — xAI launches the grok-voice-think-fast-1.0 model via the xAI Console API. The model claims first place on the Tau Voice Bench, with integrated reasoning and no added latency. It is already deployed in production at Starlink for customer support. The architecture is unified, separate from the Grok STT/TTS APIs announced in April.

🔗 Grok Voice Think Fast 1.0

Grok Imagine — improved lip sync

25 April — Grok Imagine announces improved lip sync and audio quality for all image-to-video generations.

🔗 Grok Imagine lip sync


Media generation and voice agents

Kling AI 4K — native upscaling from low resolution

24 April — Kling AI launches Kling 4K, a native 4K image upscaling feature from low-resolution sources. The announcement summarized as “Blurry in. 4K out.” generated 5.82 million views. This feature is distinct from Kling Video 3.0.

🔗 Kling AI 4K

Runway integrates GPT Image 2

24 April — Runway integrates OpenAI’s GPT Image 2 into its video creation platform.

🔗 Runway × GPT Image 2

ElevenLabs × Customers Bank — banking voice agents

24 April — ElevenLabs announces a deployment of ElevenAgents at Customers Bank (USD 25 billion in assets). Three agents are deployed: 24/7 customer support, onboarding for new customers, and real-time coaching for advisors.

🔗 ElevenLabs × Customers Bank

ElevenLabs — Ambassador Program

April 23 — ElevenLabs is opening applications for its ambassador program, which has two tiers: Community Builders and Ambassadors, with credits, goodies, and early access to new features. The announcement generated 116,000 views.

🔗 ElevenLabs Ambassador Program


Anthropic research: safety and agents

Election safeguards — evaluation results

April 24 — Ahead of the 2026 US midterms, Anthropic publishes an update on its election safeguards. Claude Opus 4.7 and Sonnet 4.6 score 95% and 96% respectively in evaluations measuring political response balance.

ModelPolitical complianceRefusal of influence operationsWeb search enabled
Opus 4.7100%94%92%
Sonnet 4.699.8%90%95%

A TurboVote banner (a nonpartisan Democracy Works resource) will be displayed on Claude.ai to direct users to reliable information on voting in the 2026 midterms.

🔗 Election safeguards update — Anthropic

Project Deal — Claude agents as negotiators

April 24 — Anthropic publishes the results of Project Deal, an internal experiment on AI agents in a Craigslist-like marketplace. For one week, Claude agents represented San Francisco office employees buying and selling items among colleagues. In total, 186 deals were completed across four parallel model configurations.

MetricValue
Duration1 week
Configurations4 (all-Opus 4.7, all-Haiku, 2 mixes)
Deals completed186
Opus vs Haiku advantage+2 deals on average, higher prices
Effect of aggressive instructionsNot statistically significant

“New Anthropic research: Project Deal. We created a marketplace for employees in our San Francisco office—like Craigslist—where Claude agents negotiated deals on their behalf.” — @AnthropicAI on X

The most notable finding: aggressive instructions (“negotiate hard”) had no statistically significant effect on the results — not because of poor instruction following, but because of the market’s own constraints.

🔗 Project Deal — Anthropic


What this means

April 25 illustrates a rapid consolidation around a few major trends. On frontier models, DeepSeek-V4 and GPT-5.5 establish a new baseline for freely accessible capabilities: one million context tokens is no longer a premium differentiator. The arrival of DeepSeek-V4-Pro for free on NVIDIA Blackwell, combined with the -75% API promotion, signals direct price competition with closed models.

On the developer tools side, the expansion of agentic autonomy is taking shape on several fronts at once — Claude Code v2.1.119, Codex Auto-review, Inline Agent Mode in JetBrains. These updates converge on the same goal: reducing human interruptions in long pipelines, while maintaining control points for risky operations. The legal framework question for agents acting on our behalf, raised by Project Deal, takes on particular resonance in this context.

The Cohere + Aleph Alpha merger with USD 600 million from Schwarz Group is the most structural signal for European sovereign AI. It creates a transatlantic player positioned on governments and regulated sectors, with dedicated cloud infrastructure (STACKIT), in a market estimated at USD 600 billion. The Anthropic + NEC partnership in parallel shows that the same sovereignty logic is extending to Asia.


Sources

This document has been translated from the fr version into the en language using the gpt-5.4-mini model. For more information about the translation process, see https://gitlab.com/jls42/ai-powered-markdown-translator