Tag

Token Efficiency

All articles tagged with #token efficiency

Opus 5 shifts focus to token efficiency over a capability leap
technology1 month ago

Opus 5 shifts focus to token efficiency over a capability leap

Anthropic’s Opus 5 delivers incremental performance gains for coding and related tasks while prioritizing lower token costs, charging about $5 per million input tokens and $25 per million output tokens. It rivals Fable on many benchmarks and is cheaper than Opus 4.x, with broader availability than Mythos, but it doesn’t train as aggressively on cybersecurity tasks and lags Mythos in vulnerability exploitation. The takeaway is cost-driven progress rather than a dramatic leap in capability, as developers explore model routers and open-weight options to minimize compute and expense.

Google doubles down on cheaper Gemini models to outpace rivals
technology1 month ago

Google doubles down on cheaper Gemini models to outpace rivals

Google unveils three new Gemini models—3.5 Flash Cyber, 3.6 Flash, and 3.5 Flash-Lite—expanding pricing and performance options: Cyber targets vulnerability detection and will be limited to governments and trusted partners, 3.6 Flash trims token use by up to 17% and lowers costs for higher-volume workloads, and 3.5 Flash-Lite offers fast, low-cost performance for large AI-agent systems. The rollout positions Google against Anthropic and Chinese rivals while advancing a broader Gemini roadmap and cost-cutting hardware efforts ahead of Alphabet’s earnings.

OpenAI touts 54% token efficiency boost for GPT-5.6 Sol amid restricted rollout and regulatory talks
technology1 month ago

OpenAI touts 54% token efficiency boost for GPT-5.6 Sol amid restricted rollout and regulatory talks

OpenAI CEO Sam Altman says GPT-5.6 Sol is 54% more token-efficient on agentic coding tasks as OpenAI rolls out GPT-5.6 Sol, Terra, and Luna in a limited launch after coordinating with US officials; he described a collaborative approval process, signaled ongoing discussions about a possible government stake, and stressed global access with safety-focused regulation.

Engram banks $98M to slash AI costs with memory-first models
technology2 months ago

Engram banks $98M to slash AI costs with memory-first models

Engram, a memory-focused AI startup, raised $98M from major investors including General Catalyst, Kleiner Perkins, Sequoia and OpenAI cofounder Karpathy, claiming its 'learned memory' tech can recall company-specific context to deliver cheaper, smarter outputs—reportedly matching or beating frontier labs with up to 100x fewer tokens—and plans to deploy the funds to grow compute and talent for enterprise clients like Microsoft, Notion and Harvey.