xAI opens Grok 3 and Grok 3-mini via API — $0.50/1M output tokens, 1/7th the price of Gemini 2.5 Flash thinking
·AI News (Buttondown)
xAI flipped Grok 3 and a new Grok 3-mini variant into general API availability via docs.x.ai. Grok 3-mini is priced at $0.50 per million output tokens — roughly one-seventh the cost of Gemini 2.5 Flash thinking — while claiming parity with much larger frontier models on reasoning traces. Developers comparing Grok 3-mini against Gemini 2.5 Pro and Claude 3.7 Sonnet report competitive tool-use performance, though with aggressive tool-call tendencies. Grok 3 had been available through the consumer X app for months without API access; the API switch is the first time third-party apps can integrate the model at scale.
xaigrokapipricinggemini
Why it matters
Grok 3-mini at $0.50 per million output tokens is the most aggressive frontier-class API pricing since DeepSeek V3 — it undercuts OpenAI's gpt-5.4-mini and Gemini 2.5 Flash thinking by multiples. If the tool-use quality holds up in production, Grok becomes the default choice for high-volume agentic workloads where per-request cost dominates model quality. Expect Anthropic and OpenAI to respond with new mini tiers or deeper volume discounts, and expect enterprise agentic products (call center bots, coding agents, retrieval pipelines) to begin A/B-testing Grok in Q2.
Impact scorecard
7.4/10
Stakes
8.0
Novelty
7.5
Authority
7.5
Coverage
6.5
Concreteness
9.0
Social
8.0
FUD risk
3.0
Coverage15 outlets · 2 tier-1
AI News, The Verge, TechCrunch, VentureBeat, Ars Technica
API release confirmed via docs.x.ai and AI News aggregation. $0.50/M price figure and 1/7 Gemini comparison are AI News editorial claims — verifiable once independent developers benchmark and publish. Treat cost comparison as directional pending independent benchmark; API availability itself is confirmed.
Kronos (AAAI 2026 accepted, arxiv 2508.02739) is the first open-source foundation model pre-trained on financial candlestick (K-line) sequences. A specialized tokenizer quantizes multi-dimensional OHLCV data into hierarchical discrete tokens; a decoder-only autoregressive transformer is pre-trained on 12B (12 billion) K-line records from 45 global exchanges. Results against the leading time-series foundation model (TSFM) and best non-pretrained baseline: 93% higher RankIC on price-series forecasting over TSFM and 87% over the non-pretrained baseline; 9% lower MAE on volatility forecasting; 22% improvement in generative fidelity for synthetic K-line sequences. Model, weights, and demo are open on GitHub (shiyu-coder/Kronos) — repo is currently GitHub-trending.
Google Research published Simula in Transactions on Machine Learning Research (April 16, 2026): a framework that reframes synthetic data generation as mechanism design, using reasoning-driven construction rather than sample-level optimization. The team (Tim R. Davidson, Benoit Seguin, Enrico Bacis, Cesar Ilharco, Hamza Harkous) generated datasets of up to 512K (512,000) data points across five domains — cybersecurity (CTI-MCQ, CTI-RCM), legal reasoning (LEXam), math (GSM8k), and multilingual knowledge (Global MMLU). Results show 'better data scales better': a 10% accuracy gain on math reasoning using Gemini 2.5 Flash as teacher and Gemma-3 4B as student. The four-step recipe is global diversification → local diversification → complexification → quality checks. Complexification helped math but hurt legal reasoning — the paper warns mechanism design is domain-dependent.
coleam00/Archon is a TypeScript open-source workflow harness that makes AI coding deterministic and repeatable through YAML-defined development processes. Hit 18.8k GitHub stars and is trending weekly. Latest release v0.3.6 on April 12, 2026 with 1,265 commits on dev branch. It ships 17 default workflows covering issue fixes, feature development, PR reviews, and refactoring. Core features: isolated execution (each run gets its own git worktree for parallel conflict-free processing), composable workflows (mix deterministic nodes like bash/tests/git with AI-powered steps like planning/code-gen/review), multi-platform (CLI, Web UI, Slack, Telegram, Discord, GitHub webhooks), and human gates (interactive approval steps). MIT licensed, requires Bun + Claude Code + GitHub CLI.