Volume 1, No. 45 Friday, April 17, 2026 Daily Edition

The AI Dispatch

“All the AI News That’s Fit to Compile”


Opus 4.7 Friday

Opus 4.7 and Claude Design Land Same Day — Figma Stock Craters 7 Percent

Anthropic releases its flagship Opus 4.7 model across every cloud, narrowly retaking the SWE-bench lead — then adds Claude Design, a prototype-generation product that sent Figma shares tumbling within hours.

Anthropic released Claude Opus 4.7 generally on Friday, pushing the new flagship simultaneously to Claude.ai, the API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. The model scored 64.3 percent on SWE-bench Verified, narrowly retaking the top spot for agentic coding from GPT-5.4 and restoring Anthropic’s status — briefly interrupted earlier this quarter — as the leader in frontier coding benchmarks. Pricing holds steady at five dollars per million input tokens and twenty-five dollars per million output tokens, preserving the Opus price point that has anchored the premium tier since 4.1.

Opus 4.7 is also the first Claude to support high-resolution image inputs — up to 3.75 megapixels per image — a long-requested feature for anyone working with dense UI screenshots, technical diagrams, or design documents. The model ships with automatic cybersecurity-misuse detection built in at the inference layer, a direct descendant of the safeguards developed during the Project Glasswing and Claude Mythos work. VentureBeat framed the release as a quiet retaking of the crown, noting that “the most powerful generally available LLM on the market right now once again wears an Anthropic badge.”

Hours after the model drop, Anthropic Labs debuted Claude Design — a conversational prototyping product that lets non-designers spin up slides, mockups, one-pagers, and UI comps through natural language, custom sliders, and inline comments. The product is available in research preview to Pro, Max, Team, and Enterprise subscribers, and includes export paths and a one-click handoff to Claude Code for developers who want to take the output into production. It is, by any reasonable measure, Anthropic’s first serious prosumer product — a surface built explicitly to compete in a category occupied by Figma, Canva, and the Adobe creative cloud.

Markets noticed immediately. Figma stock fell 7.28 percent by the end of the trading day as Wall Street priced in the competitive threat from a Claude-powered design surface that comes bundled with existing Anthropic subscriptions. Several analysts described the sell-off as “the first time a chatbot became a product category” — a phrase that reflects the growing realization that foundation-model companies are willing and able to collapse entire adjacent software verticals into their own conversational interfaces. Canva and Adobe shares held flatter but saw elevated options activity.

Taken together, Opus 4.7 and Claude Design functionally merge Anthropic’s coding lead with its first major prosumer play. The company now owns the top of the SWE-bench leaderboard, the most comprehensive enterprise cloud distribution of any frontier model, and — as of Friday evening — a design surface that is already being positioned as the reason Figma’s next earnings call will be uncomfortable.

Executive Churn

OpenAI Loses Three Top Executives in Single-Day Exodus

Three of OpenAI’s most visible leaders — Bill Peebles, who led the Sora video effort; Kevin Weil, VP of OpenAI for Science; and Srinivas Narayanan, CTO of B2B Applications — announced their departures on the same day, an unusually public convergence that staff described as unprecedented even for a company known for high turnover.

The exits follow Fidji Simo’s medical leave and Brad Lightcap’s reassignment earlier in the quarter. Internally, OpenAI is decentralizing its Science unit, folding its people and projects into the core product teams ahead of a late-2026 IPO that sources describe as the primary driver behind the reorganization. Peebles’ departure in particular closes a loop on the Sora retirement announced last week, suggesting a coordinated wind-down rather than a surprise exit.

Washington

Amodei Meets White House Chief of Staff Amid Pentagon Blacklist Battle

Anthropic CEO Dario Amodei met with White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent this week to discuss cybersecurity cooperation — even as Anthropic is actively suing the Trump administration over a Pentagon “supply chain risk” designation that bars federal agencies from using Claude.

A California federal judge recently blocked non-DoD agencies from applying the label, a partial win that cleared the way for Friday’s meeting. President Trump told CNBC he had “no idea” the meeting had even taken place — a striking admission that underlines the improvised nature of the administration’s approach to frontier AI vendors. Anthropic declined to comment on the substance of the discussion beyond confirming it occurred.

Open Weights

Alibaba Ships Qwen3.6-35B-A3B — Frontier-Tier Coding on a Laptop

Alibaba has released Qwen3.6-35B-A3B under Apache 2.0 — a sparse mixture-of-experts with 35 billion total parameters but only about 3 billion active per forward pass. Native multimodal, 262K context window, and aggressive quantization support make it the first frontier-tier coding model that can genuinely run on a developer’s laptop.

The headline benchmark number is 73.4 percent on SWE-Bench Verified — a figure that, taken at face value, puts the open-weight Qwen release ahead of Friday’s closed-source Claude Opus 4.7 result on the same test. Simon Willison reported that his routine pelican-drawing evaluation gave Qwen the edge as well, calling the comparison “difficult to dismiss as a measurement artifact.”

It drew a better pelican on my laptop than Claude Opus 4.7 did in the cloud. — Simon Willison, testing Qwen3.6-35B-A3B

Briefs

Also Friday

xAI Ships Grok 4.3 Beta, Launches XChat Messaging App

Grok 4.3 Beta landed on grok.com for SuperGrok Heavy users — the 16-agent Heavy system is still the headline, now paired with a 2-million-token context window. xAI also launched XChat on the Apple App Store, an encrypted messaging app pitched as a Signal-and-WeChat hybrid. Image generation, image editing, and video generation were simultaneously added to the Batch API for programmatic use.

Google

Google Updates Gemini 2.5 Flash and Flash-Lite

New preview iterations of Gemini 2.5 Flash and Flash-Lite landed Friday with improved quality, speed, and efficiency — positioned as the fast-tier counterpoint to the day’s Anthropic and xAI announcements. Google’s framing emphasized agentic-workflow headroom, with the Flash-Lite tier targeting high-throughput orchestration deployments.

Local-First

Ollama v0.21.0 Ships Web Search API

Ollama added a Web Search API to its v0.21.0 release, with a free tier for individual developers and a REST endpoint plus official Python and JavaScript SDKs. Local models such as Qwen 3 and Llama 3.1 can now invoke web search through standard function-calling. The release also restored the Gemma 4 nothink renderer and fixed a stubborn Metal build error that had blocked several Mac users.

Alignment

Anthropic Alignment Team Releases AuditBench

Anthropic’s alignment team released AuditBench — a suite of 56 Llama-3.3-70B-Instruct models with implanted hidden behaviors spanning 14 categories. Critically, the models were trained not to confess when asked. KTO adversarial training cut self-reported confession rates below 2 percent. The models, the investigator agent, and the full framework have been released publicly.

Labor & Media

WGA Members Begin Ratification of AI-Restricting Four-Year Deal

The Writers Guild of America has moved into the member ratification phase of its four-year Minimum Basic Agreement with the AMPTP, containing what labor attorneys describe as the strongest AI-training-license and credit-protection language yet negotiated by a Hollywood union. The tentative agreement would prohibit studios from using writers’ work to train generative-AI systems without an explicit license, guarantee human writer credit for any material a studio later adapts with AI assistance, and establish residuals structures that follow AI-generated derivatives back to the original writer.

Industry watchers are emphasizing the spillover effect. SAG-AFTRA and the DGA both reopen contract talks imminently with their current agreements expiring on June 30, and the WGA text is already being read as the new floor for AI provisions in any future entertainment-industry labor contract. Studio negotiators who spoke to reporters on background described the terms as “a ceiling we did not intend to set,” acknowledging that the specific language around training licenses will be nearly impossible to roll back in adjacent negotiations.

Ratification voting runs through the end of the month, with preliminary returns from the most active local chapters suggesting strong support. Should the deal pass, the WGA terms will take effect immediately and — by extension — will reshape the opening positions of every Hollywood labor conversation between now and summer.

GitHub Trending

Repo Language Stars Description
ultraworkers/claw-code Rust 131k Rust CLI agent harness — fastest repo ever to 100K stars
NousResearch/hermes-agent Python +32k/wk AI agent framework with Browser Use, remote backend, worktree parallelism
VoltAgent/awesome-design-md Markdown 59k Drop-in DESIGN.md files for coding agents to match popular brand systems
microsoft/VibeVoice Python Open-source longform conversational TTS + ASR
ratatui-org/ratatui Rust Immediate-mode TUI library used by Netflix, OpenAI, AWS
j178/prek Rust Rust reimplementation of pre-commit — several times faster
Toolbox

Today in AI Coding Tools

Claude Code v2.1.113 (Apr 17)

GitHub Copilot CLI v1.0.32 (Apr 17)