Token Cost Radar

1How cheap can routine AI work get? Anthropic cuts Haiku token prices by 90%

Anthropic introduced Claude Haiku 5.5 on October 7, pricing requests with prompts up to 100,000 tokens at $0.10 per million input tokens and $0.50 per million output tokens. Requests above that threshold cost $0.50 input and $2.50 output. Cache reads cost $0.01 per million tokens in the lower tier. Anthropic says approximately 90% of requests to its previous Haiku model fell below the threshold and estimates that Haiku 5.5 costs about 75% less per completed workload after accounting for changes in token consumption. The model targets classification, summarization, database queries, compaction, and routine agent substeps. It is available through Anthropic's platform and major cloud providers.

2Sonnet's cache-read price falls 50%, changing the economics of coding agents

VentureBeat reports that Anthropic's October 7 announcement also cuts Claude Sonnet 5.5 cached-input pricing from $0.20 to $0.10 per million tokens. Anthropic estimates that the change reduces costs by approximately 20% for typical agentic workloads because cached context represents a substantial share of repeated model input. The company is also introducing monthly API credits for eligible Max and Team subscribers. The practical implication is that existing applications can become cheaper without changing models, shortening prompts, or sacrificing output quality. Actual savings depend on cache-hit rates and workload composition.

3GPT-4-class inference now costs 99.8% less than it did three years ago

Gravity's pricing history, refreshed October 8, finds that the cheapest API model meeting its original GPT-4 capability threshold costs $0.0575 per million blended tokens, compared with $37.50 in October 2023. That represents approximately a 652-fold reduction. Its Gemini 2.5 Pro capability tier has fallen 94% over the past twelve months, from $2.41 to $0.152 per million blended tokens. The analysis reconstructs historical prices from LiteLLM data and uses capability thresholds rather than comparing identical models. These are cheapest-available capability-tier prices, not average enterprise invoices or guarantees of equivalent performance on every task.

4Why are enterprise AI bills still climbing when tokens keep getting cheaper?

TechNewsWorld reports October 7 that falling inference prices have not translated into proportionately lower enterprise AI spending. Industry sources interviewed by the publication point to increasing agent adoption, longer execution chains, expanding context, and the movement from experimentation into production. The article cites claims of substantial enterprise budget overruns, although those figures should be treated as attributed industry estimates rather than independently established market-wide measurements. The underlying FinOps problem is clear: a lower unit price cannot control an invoice when the number of billable units grows faster.

Agent in a Bottle: Can LLM Agents Turn Their Capabilities Into Cheap, Scalable Artifacts?

Submitted October 6, this paper introduces BOTTLED, a benchmark testing whether agents can transform repeated inference into reusable programs or specialized smaller models. Bottling means converting a general model's capabilities into cheaper task-specific artifacts that can be reused across many requests. Across ten models and three tasks, 48 of 60 bottling runs performed below the lower confidence bound of their originating model's zero-shot performance, demonstrating that the process remains difficult. Nevertheless, on query-product relevance classification, a solution produced using Opus 5 retained approximately 82% of its original macro-F1 at roughly 657 times lower reported cost. The result is task-specific and involves a quality tradeoff, but it demonstrates substantial potential for amortizing inference.

Why it matters: Repeatedly paying a frontier model to perform the same narrow operation may be economically inferior to investing once in a reusable implementation. The relevant calculation becomes development cost plus marginal execution cost, divided across the expected workload.

UNREAL: Unifying Retrieval and Long-Context with a Single Model

Submitted October 6, UNREAL introduces a model-native evidence-selection mechanism that operates across corpus retrieval and long-context inference. It adds fewer than 500,000 trainable parameters while leaving the underlying language model unchanged. On a 3-billion-token Wikipedia index containing 21 million chunks, the authors report improved retrieval recall over evaluated baselines. In long-context experiments, selecting relevant evidence before generation improved NoLiMa accuracy from 1.0% to 24.83% at 128,000 tokens and improved LV-Eval F1 from 49.97% to 54.66% at 256,000 tokens. The authors also report lower computation and time-to-first-token than full-context inference beginning around 32,000 tokens.

Why it matters: This provides evidence that reducing context can improve accuracy as well as efficiency when irrelevant material is removed intelligently. More available information does not require more resident information.

APEX: Speculate Smarter, Not Deeper

Submitted October 6, APEX introduces adaptive speculative decoding that selects a drafting strategy per request and adjusts draft length during generation. Integrated into vLLM and evaluated using Qwen3-8B across six workloads, the system achieved up to 5.24 times the speed of ordinary autoregressive decoding. Its throughput-oriented configuration averaged 4.27 times acceleration, while a more conservative configuration achieved 3.27 times acceleration with a 41% relative reduction in wasted draft-token percentage compared with fixed-depth n-gram speculation. The reported speedups are experimental serving results, not direct reductions in commercial API prices.

Why it matters: Inference efficiency depends on useful accepted computation rather than the volume of speculative work performed. Adaptive drafting can reduce wasted processing while retaining the target model's verification procedure.

TokenCast: Forecasting Token Consumption During LLM Agent Execution

This September 28 preprint remains particularly relevant to the current enterprise budgeting discussion. TokenCast predicts total agent token consumption while execution is underway, incorporating completed steps and the additional input cost created by accumulated context. Across four task suites and six agent models, the authors report an average 14.5% reduction in forecasting error relative to their strongest comparator across 96 evaluated combinations. In offline budget-control replay, TokenCast used 21.3% fewer tokens than a fixed-budget policy at matched trace completion. Forecast updates required no additional model calls and averaged 32.8 milliseconds per run on SWE-bench Verified.

Why it matters: Agent spending cannot always be estimated accurately before execution because the workflow changes as tools return results. Continuously updated forecasts offer a practical basis for enforcing budgets without terminating useful work prematurely.