Weekly AI Builder Briefing — April 29, 2026
GPT-5.5 vs DeepSeek V4 · Microsoft/OpenAI Breakup · Musk v. Altman · Budget Coding · OSS Tools
Support the venue by purchasing a drink here
<3 4seas
OpenAI shipped GPT-5.5 on April 23. The codename was \"Spud.\" They kept it in the 5.x family — not the GPT-6 everyone expected. Built for long-horizon agentic tasks: coding, computer use, research, data analysis across tools.
Terminal-Bench 82.7% (best ever). FrontierMath improvements. Same-tier latency as GPT-5.4 but uses fewer tokens per task — more efficient, not just more capable.
But: Tom's Guide tested GPT-5.5 head-to-head against Claude Opus 4.7 — GPT-5.5 lost in all 7 categories. Praised for speed, criticized for hallucinating rather than admitting it doesn't know something.
TERMINAL-BENCH
82.7%Best ever (Opus 4.7: 70%)
API PRICING
$5/$30per M tokens (= Opus pricing)
VS OPUS 4.7
0 / 7Tom's Guide categories won
CONTEXT
1Mtokens · Plus/Pro/Biz/Ent
DeepSeek shipped V4 on April 24 — two models under MIT license, both with 1M context:
V4-Pro: 1.6T total, 49B active. SWE-bench 80.6%. Codeforces 3,206 (highest ever by any model). The new largest open-weight model available.
V4-Flash: 284B total, 13B active. Nearly as good, 12× cheaper. The practical self-hosting target.
Novel hybrid attention (CSA + HCA) = only 27% of inference FLOPs and 10% of KV cache vs V3.2 at 1M context. Three reasoning modes: Non-think (fast), Think High (logical), Think Max (frontier). Trained on FP8+FP4 mixed precision. Validated on both Nvidia GPUs and Huawei Ascend NPUs.
", "notes": "The price story is the headline for builders: V4-Pro output at $3.48/M is 7× cheaper than Opus ($25) and GPT-5.5 ($30). V4-Flash at $0.28/M output is 89× cheaper than Opus. Legacy deepseek-chat and deepseek-reasoner models retiring July 24 — migration needed. Community reaction: strong on coding benchmarks, but Artificial Analysis notes increased hallucination rate vs V3.2. Always test before committing production workloads.", "media": { "type": "svg", "inlineSvg": "", "caption": "Both MIT license. Both open weights on HuggingFace. Legacy models retire July 24." } }, { "segment": "DeepSeek V4", "duration": "3 min", "title": "DeepSeek V4 — Fast Endpoints That Actually Work", "bodyHtml": "The #1 reason people don't use open-source models via API: inference speed. DeepSeek's own API runs V4-Pro at ~37 t/s — usable but not fast. Here's the cheat sheet for providers that solve this.
Together.ai delivers V4-Pro at 130 t/s — that's 3.5× faster than DeepSeek's own API and faster than most proprietary model endpoints. Lightning AI hits 82 t/s at the cheapest price ($1.42/M blended).
For V4-Flash (the daily driver), all providers deliver 81–86 t/s at an identical $0.17/M blended. At that speed and cost, there's genuinely no reason not to try it.
", "notes": "This is the slide that makes DeepSeek V4 actionable for the audience. The speed problem is SOLVED if you pick the right provider. Together.ai at 130 t/s for V4-Pro High effort is genuinely faster than Claude's API. Lightning AI is the hidden gem — 82 t/s at $1.42 blended is cheaper than even DeepSeek's own pricing. For Flash, just pick any provider — they're all fast and identically priced.", "media": { "type": "html", "html": "V4-PRO — FASTEST ENDPOINTS
V4-FLASH — ALL FAST, ALL CHEAP
On April 27, Microsoft and OpenAI renegotiated their deal. The exclusive distribution rights are gone. The AGI escape clause is dead. OpenAI is free to sell on any cloud.
Within 24 hours, Amazon announced OpenAI models on AWS Bedrock + Codex + Bedrock Managed Agents powered by OpenAI. Google Cloud is \"taking a close look\" at what's possible. Oracle already had deals in place.
The new deal: Microsoft keeps non-exclusive IP license through 2032. Azure stays \"primary cloud partner\" (first-ship rights). Revenue share capped through 2030. Microsoft stops paying revenue share TO OpenAI. OpenAI still pays 20% TO Microsoft (capped).
", "notes": "Why this matters for builders: if your company uses AWS, you'll be able to access OpenAI models through Bedrock within weeks. No more 'we can't use OpenAI because we're an AWS shop.' This is huge for enterprise adoption. The flip side: Microsoft invested $13B+ and just lost exclusive distribution. They're pivoting harder to Anthropic — already building agent offerings powered by Claude. The competitive landscape just got a lot more interesting.", "media": { "type": "svg", "inlineSvg": "", "caption": "\"We're excited to make OpenAI's models available directly to customers on Bedrock.\" — Andy Jassy" } }, { "segment": "Trial & Targets", "duration": "4 min", "title": "Musk v. Altman + OpenAI Missed Targets", "bodyHtml": "The Trial (started April 28, Oakland): Musk demanding $130B+ damages, reversion to nonprofit, removal of Altman and Brockman from the board. Jury is advisory — judge makes the final call.
Musk testified he \"came up with the idea, the name, recruited key people, provided all initial funding.\" OpenAI's lawyer: \"We are here because Mr. Musk didn't get his way at OpenAI… he quit saying they would fail. But my clients had the nerve to succeed without him.\"
The WSJ Bombshell (April 28): OpenAI missed multiple monthly revenue targets in 2026. Failed to reach 1B weekly active user goal. Subscriber defections to Anthropic and Google. CFO Sarah Friar privately worried about funding compute commitments. SoftBank dropped 10%.
Both stories threaten OpenAI's planned IPO at ~$1T valuation. The trial could force a nonprofit reversion. The revenue miss questions whether they can fund their $250B Azure commitment.
", "notes": "These are intertwined stories. The trial outcome directly affects the IPO timeline. The revenue miss raises questions about whether OpenAI's 'spend big to win big' strategy is sustainable. Anthropic gaining enterprise share is the underlying theme. Jim Cramer called the WSJ report a 'timed hit job before earnings season.' Morgan Stanley notes 'valuation dispersion' is coming for AI companies." }, { "segment": "BytePlus ModelArk", "duration": "3 min", "title": "BytePlus ModelArk — Budget Multi-Model Coding", "bodyHtml": "ByteDance's international AI platform just added GLM-5.1 (April 22). One subscription, multiple top-tier models, works with every major coding tool.
Models included: GLM-5.1, DeepSeek-V3.2, GLM-4.7, Kimi-K2.5, Seed-Code, GPT-OSS, Dola-Seed-2.0-Code. Auto mode picks the best model per task.
Tools supported: Claude Code, Codex, Cursor, Cline, Kilo Code, Roo Code, OpenCode, OpenClaw, Hermes Agent.
Data caveat: BytePlus collects your coding data (inputs + outputs) for model improvement. Anonymized and encrypted, but opt-in-by-participation. Not for sensitive client code.
LITE PLAN
$10/month · Multi-model access
PRO PLAN
$25/month · Higher quota
vs Claude Max $200/mo · vs ChatGPT Pro $100/mo
Flow Browser
Modern privacy-focused browser built on Electron/Chromium. GPLv3. What a browser looks like when you strip away all the cruft: Spaces (organized tab groups), multiple Profiles, a Command Palette for keyboard-driven navigation, Sleep Tabs for memory savings, and full Chrome extension support. No ads, no news feed, no bloat — just browsing.
T3 Code
Open-source desktop GUI for AI coding agents, by Theo (t3.gg). Wraps Codex CLI and Claude Code in a proper app with Git worktree integration, one-click PR workflows, per-turn diff viewer, and custom actions. BYOK — bring your own API key. v0.0.21 dropped April 23. Think of it as the Cursor alternative that doesn't lock you into one model.
Lightpanda — Headless Browser for AI Agents
Built from scratch in Zig. Not a Chromium fork. No graphical rendering at all — DOM + JavaScript + network only. Result: 11× faster, 9× less memory than Chrome headless. Puppeteer/Playwright compatible via Chrome DevTools Protocol. Perfect for AI agent web automation, scraping at scale, or browser-based testing. AGPL-3.0. Free tier: 10 browser hours/month.
Huashu Design — The Open-Source Claude Design
A Claude Code skill (by Chinese dev 花叔) that turns any coding agent into a designer. 7,300 GitHub stars in one week. 20 design philosophies. Produces prototypes, slides, motion graphics, infographics — all in HTML. Head-to-head tests show comparable results to Claude Design at ~1% of the token cost. Works with Claude Code, Cursor, Codex, OpenClaw, Hermes.
Lightpanda Stats
faster than Chrome
less memory
Zig · AGPL-3.0 · Puppeteer compat
Huashu Design Stats
stars in 1 week
token cost vs Claude Design
20 design philosophies · Agent-agnostic
Forgejo v15.0 released April 16 — the 100th release of this community-governed self-hosted Git forge. A fork of Gitea with non-profit governance, GPL license, and the fastest-moving feature set in self-hosted Git.
Runs on ~420MB RAM. Ships as a single Go binary. Pull requests, issues, wiki, container registry, package registry, LFS, and GitHub Actions-compatible CI (Forgejo Actions). ActivityPub federation in progress — eventually repos can talk to each other across instances.
v15.0 highlights: repo-specific access tokens, reusable workflow expansion in Actions, OpenID Connect support, ephemeral runners for secure autoscaling, improved issue filtering UI.
Fedora is replacing Pagure with Forgejo. Most 2026 guides recommend Forgejo over Gitea due to community governance and GPL licensing. Replaces GitHub at ~€12/mo on a Hetzner VPS.
", "notes": "forgejo.org. The recommendation for the audience: if you're paying for GitHub Team ($4/user/month), a single Forgejo instance on a €7.50 Hetzner box gives you unlimited users, unlimited private repos, CI/CD, container registry, and full data sovereignty. The GitHub Actions compatibility means your existing workflows mostly just work. Docker one-liner: docker run -d -p 3000:3000 codeberg/forgejo:15. Pairs perfectly with Tailscale for secure access without opening ports.", "media": { "type": "svg", "inlineSvg": "", "caption": "docker run -d -p 3000:3000 codeberg/forgejo:15" } }, { "segment": "Wild Stories", "duration": "3 min", "title": "Wild Stories This Week", "bodyHtml": "SOUTH AFRICA WITHDREW ITS AI POLICY
The country's first draft national AI policy was pulled after it was discovered to contain fictitious sources that appeared to be AI-generated hallucinations. A government using AI to write AI policy, and the AI made up its citations. You cannot make this up.
YELP AI ASSISTANT COMPLETES BOOKINGS
Yelp expanded its AI assistant to handle the full journey: discover a restaurant, book a reservation, order food, schedule a service — all within a single conversation. Discovery + transaction in one interface. This is where agentic AI hits real revenue.
AMAZON \"JOIN THE CHAT\"
Amazon's new feature lets you ask questions about products and receive AI-powered audio responses. Shopping by conversation. The product page is becoming a dialog, not a static listing.
GEOFFREY HINTON AT UN
The \"godfather of AI\" at the Digital World Conference: \"If you ever went out with a car that had no brake… you are in trouble. But you're in even more trouble if there's no steering wheel.\"
TRY DEEPSEEK V4-FLASH
$0.14/$0.28 per M tokens at 81+ t/s. Use Novita or SiliconFlow. Swap in deepseek-v4-flash and compare against whatever you're paying for now.
INSTALL T3 CODE
If you use Claude Code or Codex from the terminal, T3 Code gives you a proper GUI with git worktrees, PR workflows, and diff views. Free, open-source, actively maintained.
SPIN UP FORGEJO
One Docker command. GitHub Actions-compatible CI. Unlimited everything. €12/mo instead of $4/user/month on GitHub Team.
CHECK YOUR OPENAI DEPENDENCY
The Microsoft breakup, revenue miss, and trial all create uncertainty. If you're building on OpenAI exclusively, this is the week to test alternatives. DeepSeek V4, Kimi K2.6, and Qwen 3.6 are all viable.
ALWAYS VERIFY AI CITATIONS
South Africa's government didn't. Their entire AI policy got pulled. If a government can get burned, so can you.
Groups of 3–4 · Pick 2–3 questions · 10 minutes
GPT-5.5 lost to Claude Opus 4.7 in all 7 head-to-head categories. Does that match your experience? Which model are you actually using day-to-day?
DeepSeek V4-Flash at $0.28/M output tokens is 89× cheaper than Opus. At what price point does \"good enough\" beat \"best\"?
OpenAI on AWS Bedrock changes everything for enterprise. Does this make you more or less likely to build on OpenAI?
Which of the OSS tools (Flow, T3 Code, Lightpanda, Huashu Design, Forgejo) would you actually try this week? Why?
If the Musk trial forces OpenAI back to nonprofit and the revenue miss tanks the IPO — what happens to the AI industry?