Brought to you by:
Top stories today:
- GPT-5.5 debuts as stronger than Opus 4.7, weaker than Mythos
- DeepSeek releases V4 Pro, V4 Flash open-weight models in preview
- Anthropic fixes Claude Code issues tied to reasoning, caching
- Meta, Microsoft plan cuts, buyouts affecting ~23,000 jobs
- Cohere, Aleph Alpha agree merger at ~$20B to build āsovereignā AI
- Cognition said to eye raising hundreds of millions at $25B valuation
0. Data and calendar
All values as of 6 AM ET / 3 AM PT, other than S&P500 and NASDAQ close (4 PM ET / 1 PM PT).
All times are ET.
Listen to our AI-generated podcast summarizing todayās newsletter (beware of hallucinations):
1. GPT-5.5 debuts as stronger than Opus 4.7, weaker than Mythos
GPT-5.5 topped the Artificial Analysis Intelligence Index (of public models), overtaking Claude Opus 4.7 and Gemini 3.1 Pro:
GPT-5.5 will power OpenAIās upcoming āsuper app,ā including chatbot, coding tool, and web browser.
Priced at $5/$30 per 1M input/output tokens (~2Ć GPT-5.4), vs. Opus 4.7 ($5/$25), Gemini 3.1 Pro ($2/$12 ā¤200K), and Kimi K2.6 ($0.95/$4).
GPT-5.5 Pro costs $30/$180 per 1M tokens.
Gains are strongest in agentic coding, compute use, and early scientific research, OpenAI said.
Rolling out to Plus, Pro, Business, and Enterprise users across ChatGPT and Codex.
GPT-5.5 Pro will be available to Pro, Business, and Enterprise users in ChatGPT.
Anthropicās Mythos is taking on GPT-5.5 across benchmarks
Similarly to GPT-5.4, GPT-5.5 (xhigh) has an 86% hallucination rate vs. Opus 4.7 (36%) and Gemini 3.1 Pro (50%)
Watch the on-demand webinar to see how Slackbot lives in the flow of work, already understanding your conversations, files, and peopleādelivering answers and actions that are relevant and personalized to you.
Learn from real Slack customers, like Engine and Asymbl, as they share how Slackbot finds answers, analyzes documents, schedules meetings, creates content, and orchestrates specialized agents to execute complex workflows.
See contextual intelligence in action as Slackbot moves work forward by respecting your permissions and working with the tools you trust.
*Sponsored.
3. DeepSeek releases V4 Pro, V4 Flash open-weight models in preview
V4 Pro topped open-weight models on the GDPval-AA leaderboard:
V4 Pro is a significant upgrade over V3.2 in agentic capabilities.
1.6T parameters (49B active), DeepSeekās largest model.
V4 Flash has 284B parameters (13B active).
MIT licensed.
Text-only input/output, similar to V3.2.
V4 Pro costs $1.74/$3.48 per 1M input/output tokens vs. GLM-5.1 ($1.4/$4.4) and MiniMax-M2.7 ($0.3/$1.2).
V4 Flash costs $0.14/$0.28 per 1M tokens.
Available via DeepSeekās API initially.
Ask Attio anything: prep for calls, update records, spot deals at risk
Over 4,000 startups, including Granola, Wispr Flow, and Lightdash, already use Attio.
*Sponsored.
5. Anthropic fixes Claude Code issues tied to reasoning, caching, verbosity
Default reasoning set to āxhighā for Opus 4.7 and āhighā for other models to boost intelligence:
Overlapping experiments and a hard-to-reproduce bug delayed detection; fixed on Apr. 10.
Apr. 20: Anthropic reverted a prompt change that reduced verbosity but hurt performance.
Anthropic is adding tighter controls on system prompt changes.
The company will also have more staff use the exact public build of Claude Code.
Need domain experts for AI training work? Athyna Intelligence connects AI labs with financial analysts, economists, and quant researchers from Latin America who can evaluate models on:
Financial modeling & valuation
Investment analysis & portfolio theory
Risk assessment & derivatives pricing
Same US timezone. Ready in days. 40ā60% cost savings.
*Disclaimer: We have equity in Athyna.
7. Meta, Microsoft plan cuts, buyouts affecting ~23,000 jobs as Meta targets ~10% layoffs
Meta plans to cut ~8K jobs starting May 20 as part of streamlining efforts:
