Topic
LLM provider feature reusing computed key-value tensors for repeated prompt prefixes, cutting latency and cost until cache expiry.
Anthropic switched on idle compaction in Claude Code around v2.1.286, replacing long idle conversations with a summary, a dev.to analysis reports. Version 2.1.290 adds idleCompaction: false, a choice that falls on anyone keeping large sessions open past an hour.
Reality
A dev.to test measured Claude Code 2.1.289's /compact at $0.056 to $0.059 on a warm 136k-token session and $0.53 to $0.54 after the cache expired. The default cache lifetime is five minutes, so a longer pause sets the price.
Reality
Ollama's OpenAI-compatible endpoint reported 3,875 prompt tokens per second where the real rate was 125, a dev.to author found. Any benchmark that repeats a prompt and reads that field overstates prefill speed by the share served from cache.
Reality
Anthropic launched Claude Haiku 5.5 at an average price about 75% below Haiku 4.5. The saving varies widely with prompt length, so teams moving classification, support or query traffic need to price their own requests before they switch models.
Perspective Coverage
22 publishersReality
Anthropic cut the Claude Sonnet 5.5 cache-read rate from $0.20 to $0.10 per million tokens on October 7. A team saves $0.10 for each million cache-read tokens it logs, so its total bill drops by half the share those reads already held.
Reality
Amazon Bedrock serves GPT-6 Astra through two endpoints that split explicit prompt caching from invocation logging, a dev.to analysis of the model card finds. Teams have to choose per request path, and the model's end-of-life, no sooner than September 8, 2027, sets how long that choice must hold.
Reality
The Neuron gave it a browser, a Mac, a broken Blender install and an hour. What changed was how rarely it stopped to ask permission, which makes the next piece of work a harness problem rather than a prompt problem.
Perspective Coverage
13 publishersReality
Claude Code's /cost command now counts cache misses and, when it can, names a likely cause for the last one. Developers who leave sessions open over a break can now see how much usage a cold cache costs them.
Reality
Anthropic released Sonnet 5.5 at $2 and $10 per million input and output tokens, half the Opus 5.5 rate. How much a buyer saves by moving work down a tier depends on tokens burned per task and on cache reads priced identically on both models.
Perspective Coverage
5 publishersReality
Microsoft released metadata from 301,026 GitHub Copilot agent sessions, covering 9.3 million LLM calls in one June week. Capacity planners now have real cache and token figures to test against, though the files measure resource use only and cannot show whether the output was any good.
Reality
One developer's month of Claude Code logs shows each break past the one-hour cache lifetime rewriting 125k-135k tokens at twice the base input price. What that costs a team depends on its cache timer and how big sessions grow before a break.
Reality
Cache-augmented generation costs about what retrieval does when the corpus is roughly 10 times the tokens retrieval would send, a dev.to analysis finds. Sparse traffic breaks the rule, because each query then pays the cache-write premium and caching becomes the most expensive option.
Reality
OpenAI's GPT-6.1 Sol halves old Sol's cache-read rate to $0.10 per million tokens and requires the Responses API for tool calls. In one worked example the cut saves about 10%, and older agents only get that saving after their tool and reasoning fields are rewritten.
Reality
OpenAI's GPT-6.1 Sol halves cached-input pricing to $0.10 per million tokens and leaves standard rates at $2 and $10. Agents that resend long context collect the saving, while other buyers weigh gains shown mostly in OpenAI's own tests.
Reality
Opus 5.5 matched Opus 5 on two reasoning puzzles in The New Stack's tests at 43 to 69 percent lower cost. Both ran at default effort, medium on the new model and high on the old, so the saving a team sees depends on the effort level it pins.
Perspective Coverage
17 publishersReality
AWS Cost Anomaly Detection works from Cost Explorer data up to 24 hours old, so a dollar alarm on an agent fires after the money is spent. OpenTelemetry's GenAI spec has no cost attribute either, so teams must price each span from cache-split tokens and sum the trace tree.
Reality
OpenAI priced GPT-6 Sol at $2/$10 and Luna at $0.10/$0.50 per million tokens on September 22, half its GPT-5.6 promotional rates. Moving a job from Sol to Luna cuts its token rate by 95%, a bigger saving than the halving for any team whose work Luna can handle.
Reality
Jev's own gateway benchmark shows routing raised Opus 5 input tokens 61% on a Claude Code feature task, where the gateway can only hint at tools. Any saving depends on the task and on how many tools Claude Code sends the router each turn.
Reality
Every off-peak rate sits above the old flat price, and Pro cache hits jumped roughly 6x. Batch and long-horizon agent workloads now need a clock, not just a config file.
Reality
A 20% input and 33% output cut on GPT-5.6 Sol comes with a November 21 floor, while per-request regional routing makes data residency cheaper than plain global processing was in July.
Reality
Product · September 22, 2026 · Confirmed5 publishers
Science · September 23, 2026 · Confirmed2 publishers
Product · September 23, 2026 · One report1 publisher
Build · September 2, 2026 · One report1 publisher