Cost & efficiency
AI Gateway now shows where a session’s cache and spending went.
Inspect session-level routing and cache reuse before increasing the model budget.
NO FEED / Topics
Cost & efficiency
Pricing, caching, usage and the cost of an accepted result.
22 results
Cost & efficiency
Cheap subagents need a result and a stop.
A developer chose cheaper models for delegated tasks and still saw most usage attributed to subagents. Measure the accepted result, including handoffs, retries and review.
5 min readCost & efficiency
Cursor’s new bundle shares Grok usage. Claude and GPT are billed separately.
Keep the current plan until the model-level budget and permanent account changes are clear.
Cost & efficiency
Haiku 5.5’s lowest rate ends above 100K prompt tokens.
Evaluate one narrow subagent role and measure completed-task cost on both sides of 100K.
Cost & efficiency
GPT-6.1 Sol is out at Sonnet 5.5's per-token price, with cached input at half Anthropic's rate. For agent loops that re-read context, it goes on your eval this week.
Run it against your own task set before you move traffic. The price is verified; the capability claims are OpenAI's own runs.
Cost & efficiency
Anthropic and OpenAI both cut prices on the same Tuesday, and Anthropic's cut arrives attached to three migration errors.
Worth moving to, but read the migration notes before you change the model string. The three 400s are the whole of the work, and each has a named replacement.
Cost & efficiency
GPT-6 prompt caching gets a 30-minute reuse window, effort changes that survive the cache, and a miss diagnostic
Anyone running a long-lived agent loop on GPT-6 whose input bill is mostly re-sent context.
Cost & efficiency
GPT-6 Sol and Luna ship at half their predecessors' prices
Anyone with a Sol or Luna line in a production bill, and anyone who priced an agent out at the old rate.
Cost & efficiency
Grok 4.7 is rolling out in GitHub Copilot, billed at provider list pricing
Copilot Business and Enterprise admins, whose default-enablement policy turns new models on unless they opted out.
Cost & efficiency
The harness is a pricing decision
A developer choosing an agent also chooses how that agent spends. A September comparison kept the model fixed and found a large cost gap. Its evidence is useful; its limits matter before a team switches tools.
5 min readCost & efficiency
HarnessTax priced 21 model-harness pairs. The harness moved the bill further than the score.
The design and the raw chart data are public and worth an hour. They will not settle a harness migration on thirty tasks — price your own rollouts before you move.
Cost & efficiency
TypeSafe prices Jev at $0.042 per million input tokens with free output, then gates it behind a waitlist.
Join the waitlist and keep the pricing table. Hold the verdict until somebody outside the company has run it on their own workload.
Cost & efficiency
The discount is a cache hit
A cheaper cache read helps only when the client actually reuses the context. Claude Code’s September rollout of Fable 5.1 shows how a working agent can finish its task while missing the economics used to justify it.
5 min readCost & efficiency
Claude Code 2.1.260 · Fable cache + effort
Anyone on Fable 5.1 in Claude Code after 2.1.257–259
Cost & efficiency
Claude Code 2.1.260 fixes Fable 5.1 cache and effort mid-session; npm latest is 2.1.260.
Pin Claude Code ≥2.1.260 if you run Fable 5.1. On Copilot, enable Gemini 3.8 Flash and plan the 2 Oct model cutover.
Cost & efficiency
ChatGPT Plus gets the Codex and Work 5-hour cap back — Pro $100/$200 stay uncapped for now.
Plus Codex/Work: plan around a 5h rolling window again, or move Pro / credits / another harness. Claude Code on Bedrock or a gateway: stay on latest ≥2.1.239 (ideally 2.1.246); stable still lacks those fixes.
Cost & efficiency
GPT-5.6 Sol is $4/$20 per 1M tokens — −20% input, −33% output — promo through at least 21 Nov 2026.
If you already call Sol (or gpt-5.6), re-check invoices and budget alerts against $4/$20. Prefer Batch or Flex for offline work. Pin gpt-5.6-sol if you need the promo rate to stay attached to a snapshot name.
Cost & efficiency
DeepSeek peak/off-peak API rates are live as of 16 Aug 16:00 UTC.
Meter DeepSeek by UTC hour. Prefer off-peak for batch and long loops; pin Pro-0813 or Flash-0731 so a silent version cell cannot surprise you.
Cost & efficiency
DeepSeek API · peak/off-peak prices from 16 Aug 16:00 UTC
anyone metering DeepSeek as a default agent model
Cost & efficiency
goose 1.46.0 · unrolled loop, cache-safe turns, per-message cost
goose hosts metering multi-turn runs
Cost & efficiency
Claude Sonnet 5 · $2/$10 made permanent
anyone whose agents default to Sonnet 5 and who had a 50% step-up in the September forecast
Cost & efficiency
GPT-5.6 price cut · Fast mode
anyone with a high-volume batch job costed at last week's prices