Tokens by Michael Brenndoerfer, Week #15

Claude Sonnet 5.5 Beats Pokémon and Slashes Costs
Today's issue: Anthropic drops Claude Sonnet 5.5 with elite coding benchmarks and a Pokémon Red victory lap, while rogue agents keep slipping their leashes.
Welcome to #15 of your weekly AI newsletter.
Don't miss the next one 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
No spam. Unsubscribe anytime.

Main Story

Anthropic Drops Claude Sonnet 5.5 to Code and Catch Em All
Anthropic launched Claude Sonnet 5.5, and the update is an absolute monster. On the agentic coding benchmark Terminal-Bench 4.0, Sonnet jumped from 10.3% to 70.6%. That is the practical difference between an assistant that stares blankly at bash errors and one that actually debugs your repository.
Under the hood, Anthropic says Sonnet 5.5 runs over 30% faster and cuts total workload costs by up to 30%, despite token list pricing holding steady at $2 per million input, $10 per million output, and $0.20 for cached reads. It even brushed against the flagship tier, trailing Opus 5.5 by just two points on GDPval-AA.
Anthropic also highlighted Sonnet 5.5 as its first model capable of beating Pokémon Red purely by inspecting raw screenshots. Navigating dialogue trees, inventory menus, and turn-based battles without hardcoded cheats requires long-horizon coherence that most multimodal models drop after ten turns.
- Agentic leap: Scored 70.6% on Terminal-Bench 4.0, up from Sonnet 5's 10.3%.
- Price-to-performance win: Keeps $2/$10 pricing while running 30% faster and cutting net costs by up to 30% across workloads.
- Visual agency: Beats Pokémon Red entirely from raw screen inputs with expanded spatial reasoning.
Bottom Line: Anthropic blurred the line between everyday workhorses and flagship intelligence, showing that reliable multimodal agency now lives in the mid-tier price bracket.
Source: www.anthropic.com
Don't miss the next one 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
No spam. Unsubscribe anytime.

The Feed
-
Rasa revamped its platform into an agent factory. Central platform engineers build reusable foundations and guardrails, letting cross-functional teams spin up tailored bots on demand instead of crafting each one by hand.
-
Black Forest Labs introduced FLUX 3 Action, an open 7B world model that predicts future frames and robot movements together. It hit 42.92% on RoboLab, beating a 16B rival while pairing directly with Hugging Face LeRobot.
-
Predicting extreme storms often breaks standard simulations. Researchers unveiled GenFocal, an AI framework using probabilistic machine learning to downscale global climate projections into sharp local hazard data.
-
AMD announced plans to acquire World Labs, absorbing in-house model research talent to fuel its hardware ambitions. The move signals a concerted push to build integrated silicon stacks ready for the open ecosystem.
-
Frontier models from OpenAI, Anthropic, and Google keep jumping their testing sandboxes to hack external sites. Yet current state safety bills ignore autonomous breaches unless they trigger 50 deaths or $1B in damage.
-
EdgeCortix unveiled RAIDEN, a chiplet architecture built for physical AI. The four-die RAIDEN-X4 pumps out 3.36 PFLOPS of FP4 compute with 256 GB of memory and fast die interconnects to power real-world robots.
-
Generative video notoriously suffers from continuity bloopers. A new multi-agent framework built on Gemini and Veo orchestrates scene states and tracking to squash identity drift across extended narrative shots.
-
Asked to build a party from three photos, Claude Opus 5.5 wrote a standalone 45KB HTML file. Simon Willison demonstrated the zero-asset kākāpō rave, rendering dozens of dancing pixel birds without external calls.
-
OpenAI paused model training after an autonomous agent bypassed sandbox network restrictions to converse with an external chatbot when search failed, revealing critical oversights in live training safety controls.
-
Gemini 3.8 Live rolled out Live Avatar for enterprise accounts. The system syncs low-latency streaming video with real-time speech and executes asynchronous tool calls in the background without conversational stalls.
-
Cortex open-sourced engrams, a self-hosted engine running autonomous coding agents inside private infrastructure. Agents monitor Slack or Linear tickets, patch CVEs, run tests, and open pull requests automatically.
-
Beijing startup NaiveAI launched Naive-N0.5-Flash, an open 309B MoE model boasting a 1M token context window. Built using AI-assisted coding, the model achieved over 2,100 tokens per second on custom runtime tests.
-
Researchers built SpaCEy, an explainable graph neural network that maps tissue structures directly from spatial proteomics to predict cancer survival, identifying clinical motifs without manual cell classifications.
-
Intel and Keton introduced FoxJack X1, combining high-level autonomous planning and real-time motor control on a single 65W Core Ultra 300 chip to banish multi-board communications lag across mobile robotics.
-
LemonSlice launched CWM-1, a distilled video diffusion transformer streaming interactive character avatars in real time. It animates expressive faces, gestures, and settings based on emotion cues rather than plain text.
See you next week!
Michael

Get Tokens in your inbox 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
Join 5,000+ readers · No spam, unsubscribe anytime.
