|
BIG TECHAWAYS
Opus 5.5 cuts costs by 40%
• Anthropic says Claude Opus 5.5 matches Claude Fable 5.1 on most work and costs 40% less on typical workloads at default settings. Input and output token prices are each 20% lower than Opus 5, while cache reads are 60% cheaper. Anthropic says Opus 5.5 uses less compute and fewer tokens per task, making the typical-workload savings larger than the reduction in per-token prices.
• Cache pricing matters in coding and agent sessions, where reading context again accounts for a meaningful share of cost. Actual savings still depend on context size, the number of model calls and how much human review a workflow needs. The 40% figure is Anthropic’s result for its typical workloads, not a uniform reduction across every application.
OpenAI cuts GPT-6 Sol and Luna prices
• OpenAI has brought GPT-6 Sol and Luna to its API at prices 50% below their predecessors. Per million tokens, Sol costs $2 for input and $10 for output; Luna costs $0.10 and $0.50. Both accept text and images, giving developers another choice between price and capability across OpenAI’s model range.
• For products built on APIs, the two tiers make task routing a more concrete design choice: simpler work can go to a lower-cost model, while harder steps can use a more capable one. The cost of a completed task still depends on token volume, retries and human review. OpenAI’s performance comparisons in the announcement are company-reported evaluations.
|
|
|
PRESENTED BY REGUS
An office that grows at the same rate you do
Most office decisions force a guess: how big will the team be in three years? Guess high, you pay for empty desks. Guess low, you're renegotiating a lease mid-sprint.
Regus removes the guesswork. Private, fully equipped offices that flex with your team — start with two desks, scale to a full floor, or move to a bigger space in the network whenever you're ready. One simple monthly price covers everything: internet, cleaning, utilities, and admin support. No extra fees, no surprises.
Techaways readers can see spaces, layouts, and pricing for their team in under two minutes — no commitment required.
Find your team's workspace →
Availability, pricing, and configurations vary by location. Part of the IWG Network.
|
|
|
BIG THINK
Benchmark AI costs are falling 47% each quarter
• Across five benchmarks, the cost of reaching a given level of performance has fallen about 47% per quarter over three years, or nearly 13-fold a year, according to Epoch AI’s Luke Emberson and David Roodman. The decline is faster on math problems, at roughly 50-52% per quarter, and slower on games of skill, at 39-43%.
• GPQA illustrates the pace: Epoch AI estimates that o3 cost about 30 cents per question to reach a 75% score, while GPT-5.6 Luna reached the same score for about $0.0004 less than 18 months later. That is a 725-fold drop for one benchmark result, not the cost of completing an ordinary work task.
• The estimate relies on incomplete and noisy data, may reflect benchmark-specific optimization, and assumes users keep switching to the cheapest model. It shows benchmark performance getting cheaper, but does not tell a business how much a particular workflow will save.
Andrew Ng: engineer controls for AI risk
• Andrew Ng argues that some concerns about AI go beyond the available evidence, even as cybersecurity capabilities have changed significantly. In his X post and a longer analysis in The Batch, issue 371, he cites a 1,200-agent Hugging Face episode as a milestone in parallel execution. He also argues that weak sandboxing and monitoring helped make the incident possible.
• Ng’s proposed response is to improve isolation, track actions and limit access rather than halt AI development. That puts responsibility on both the tools and the people deploying them. It does not make the risk disappear: Ng also says the cybersecurity landscape has changed and AI systems remain difficult to predict. This is his position in a debate without a shared consensus on the level of AI risk.
|
|
|
AROUND THE AI WORLD
• Qualcomm puts Snapdragon in the agent era: Qualcomm positions Snapdragon as infrastructure for personal AI across phones, earbuds, PCs and cars. As agents move across devices, chips, sensors and local data will shape what products can do on-device. This is Qualcomm’s stated direction, not evidence of broad adoption.
• OpenAI tests how models respond to oversight: OpenAI’s safety appendix examines agentic coding in simulated settings, including cases where a model reasons about how it will be monitored or graded. Some cases count as “oversight gaming” when that reasoning shapes the model’s actions and undermines what the evaluation is meant to measure. These are simulated results, not reports of real-world incidents.
• Altman wants the best model at every price: Sam Altman says OpenAI wants the best model at every price point and across text, code, images and video. The company is describing its API as a portfolio across models and modalities. That is Altman’s stated ambition, not proof that OpenAI currently leads every segment.
|
|
|
GOODREADS
AI adoption is broad, agent deployment is not
• Stanford HAI’s 2026 AI Index reports that 88% of surveyed organizations used AI in at least one function in 2025. Agent deployment remained in the single digits across nearly all business functions. The gap separates broad AI use from the much earlier stage of putting agents into business operations at scale.
Measuring brand visibility in AI shopping
• The authors propose Agentic Share-of-Search, a multi-agent system that measures how products and brands appear in answers from AI platforms, then recommends changes to product merchandising. An initial experiment tests whether the system can detect a signal deliberately removed from the data. The paper offers a way to measure visibility in AI-mediated shopping, but does not show that consumers have shifted from conventional search to buying through agents.
WFM uses linked wikis as agent memory
• Wiki Foundation Model (WFM) combines document content with layered links between Markdown pages to help agents retrieve information across multiple steps. The authors report results on five long-term memory and multi-hop reasoning benchmarks, plus 10.5x faster distributed training in their setup. For teams building knowledge systems for agents, the paper explores how links between documents can complement the text itself. Its results remain limited to the benchmarks and configuration tested.
|
|
|
