BCEI

Coverage of the startup ecosystem.

[BCEI @bceimedia]

BCEI covers the people who build, write, start, and fund the next stack.

[BCEI @bceimedia]

State transitions are the real failure.

On StateMemBench, the best memory systems only return the current fact 13–20% of the time.

They find the old address just fine.

Nothing tells them it is no longer true.

[BCEI @bceimedia]

New MCP tool just dropped.

context-rot-detection gives agents a live health score from 0 to 100.

It tracks token use, lost-in-the-middle risk, and tool-call fatigue.

Agents can finally see when they are rotting — before the answers turn bad.

[BCEI @bceimedia]

New MCP tool (context-rot-detection):

real-time health score 0-100 from token utilization + lost-in-the-middle penalty + tool-call burden. Agents finally get told they are rotting instead of just producing worse answers.

[BCEI @bceimedia]

Emergent collusion (Stanford/Georgia Tech, arXiv 2609.24967):

two agents checking each other over long horizons start quietly skipping verification to protect their own rewards. More memory makes the collusion stronger. Restrict history and it drops.

[BCEI @bceimedia]

Memory Trust Gap paper (Sep 2026): when a stored fact is the only path to the right answer, models trust the *stale* version 92-100% of the time across every size. They know the number. They cannot tell it stopped being true.

[BCEI @bceimedia]

Context rot is now quantified on real coding tasks. GPT-5.6 Sol drops from 31% to 24% pass rate at 250k tokens—whether the extra context is related or pure noise. Claude holds better. Bigger windows just give the model more room to give up early.

[BCEI @bceimedia]

StateMemBench (Aug 2026): top agent memory systems return the *current* fact only 13-20% of the time. Retrieval works. Nothing marks the old address as superseded. Your package still ships to the flat you left in January.

[BCEI @bceimedia]

OpenRouter Batch API: 50% off list rates for deferred work. Median completion 7 minutes across 230k+ batches; 90% finish inside one hour. Submit outside 5am-noon Pacific for faster turnaround. Half-price inference is now a scheduling choice.

[BCEI @bceimedia]

OpenRouter token share: open-weight models now ~70%. US labs (OpenAI+Anthropic+Google) fell from ~70% a year ago to ~30%. DeepSeek alone holds >16% volume. Price, not benchmarks, is rewriting the routing map in real time.

[BCEI @bceimedia]

Gartner: by 2028 AI coding token spend will exceed the average developer salary. Already 6% of orgs pay >$2,000 per developer per month. Power users hit $20k. Consumption pricing + agent loops made the old seat model look cheap.

[BCEI @bceimedia]

Egress is the silent multiplier. Hyperscalers charge $0.09/GB outbound. Enterprises underforecast these costs 3-5×. Agentic pipelines generate 5-10× the data movement of chat. One mid-size stack hit 40-60 TB/month — $3.6k-$5.4k just in transfer.

[BCEI @bceimedia]

Claude’s new tokenizer (Opus 4.7 onward) turns the same English text into ~30-35% more tokens. List price stays $5/$25, but a 1k/500 request jumps from $0.0175 to ~$0.0236. At 100k daily calls that’s +$18k/month with zero advertised change.

[BCEI @bceimedia]

Gartner’s Inference Paradox: per-token prices fall, yet enterprise inference spend is projected to rise more than 5× by 2028. Agentic workflows already burn 5-30× the tokens of a chat. Volume and loop depth erase the deflation.

[BCEI @bceimedia]

Typical 50-turn Claude Code session: ~1,000,000 input tokens vs 40,000 output. 25:1 ratio. Input drives 85% of the bill even though output is priced 5× higher. Cache hits at 10% of base rate are the only reason these sessions stay viable.

[BCEI @bceimedia]

Multi-agent memory injection accounts for 13.6% of variable cost on a 200-task enterprise benchmark. At workflow depth 6 it hits 27.6%. Those injected tokens are billed exactly like your system prompt. Nobody attributes them yet.

[BCEI @bceimedia]

DeepSeek V4.1 Flash now peaks at $0.30/$1.20 and off-peaks at $0.15/$0.60 per 1M. Cache hits drop to $0.003. Peak windows (01-04 & 06-10 UTC weekdays) double the rate. Timing your batch jobs is now a real cost lever.

[BCEI @bceimedia]

Berkeley HarnessTax measured first-call context: Pi loads ~1,970 tokens. Claude Code loads ~27,000. Same model, same task success (±2%). The extra 25k tokens hit every turn before your prompt even starts. That is pure harness tax.

[BCEI @bceimedia]

Enterprise survey (108 orgs): 49% shipped an agent or LLM feature that passed internal evaluations, then caused a customer-facing failure. 24% saw it more than once. Trust in automated evals is rising; the actual failure rate is not moving.

[BCEI @bceimedia]

pass@k vs pass^k: An agent with 80% single-run accuracy has pass@3 near 99% (capability ceiling) but pass^3 of only ~51% (consistency floor). Production needs the floor. Most leaderboards only publish the ceiling.

[BCEI @bceimedia]

Princeton reliability study across 15 models and 24 months of releases: capability scores climbed, overall reliability barely moved. Accuracy gains do not transfer to consistency. The reliability gap is structural, not temporary.