Cost Optimization

Morph Ultra offers unlimited Kimi and DeepSeek for $200 a month, initially capped at 1,000 subscribers. Its operator says rate limits apply only to uncached prefill tokens.

Fable 5.2 leads GDPx at 16.4%, with gains concentrated in chip design and software engineering. SemiAnalysis reports average task costs rose 31% over Fable 5.1.

OpenRouter launched Auto Code, reporting Opus-level coding performance at 74% lower cost. Its routers explain decisions and adapt to model restrictions; beta reportedly cut invalid tool calls fourfold.

OpenRouter reports that agents consumed roughly 600T of its 790T October tokens. Agent cache hit rates exceeded 90%, while effective token prices fell 57–70% from January.

Vals tested seven models and found four retained at least 80% of their scores at half the cost or less. All seven scored worse overall at their lowest reasoning setting.

Runner-up packages a self-hosted setup used for thousands of daily CI jobs, with disposable VMs for isolation. Its author estimates equivalent cloud runners would cost $5,000–$9,000 per month.

Theo says T3 Code's comparisons aggregate real-world usage rather than matched tasks. Critics cite prompts, effort settings, custom tools and cache handling as factors that could affect the results.

Dagger for GitHub runs the same checks locally and after push, with results in terminals and pull requests. Dagger Cloud caches checks and bills actual usage rather than allocated machine time.

Droid 1.0 batches tool calls into programs. Factory reports up to 20% better efficiency on Terminal-Bench 4 and a Linear migration task dropping from 3.1M to 61K input tokens.

Copilot Fusion coordinates Claude and GPT models for coding work. In a GitHub-funded Artificial Analysis test, Fusion Max cost $5.59 per task versus $11.42 for Claude Code.

Arena reports a $0.40 median task cost for GPT-6.1 Terra Max, 28% below its predecessor. Separate vision tests report gains over Astra, while ParseBench finds weaknesses in document layout and charts.

Google's Logan Kilpatrick says Gemini API spend caps now have a ten-minute enforcement delay, down from more than a day. He distinguishes caps from email budget alerts and says immediate enforcement is not yet available.

Vals reports Kimi K3.1 Pro completes Terminal-Bench 4 runs roughly four times faster than K3 at a third of the cost, but regresses on legal and tax tasks. Agent Arena places it eleventh with a $1.03 median task cost.

Modal released Spindle, which runs unchanged Tinker scripts in a Modal workspace with GPU-time billing. Modal reports 20–80% lower per-token costs when training runs share GPUs.

Cua Spaces added early-access deployment in customers’ AWS and Hetzner accounts. Cua reports lower sandbox costs than E2B and provides a compatibility command for existing E2B programs.

Vals reports that Mercury 3 scores 60.9% on Vibe Code Bench and costs $1.25 per app, versus $0.38 for its predecessor. It also reports weak terminal results and crashes when output budgets are reached.

Cohere released Rerank 5 in Pro and Fast tiers, with Context Focus for selecting relevant passages. It claims up to 41% fewer downstream tokens and nearly three times Pro's throughput with Fast.

Teortaxes reports that log audits found agents repeatedly tweaking work after completion. Tests with revised Agents.md instructions reduced usage from 49M to 15.5M tokens, with a later run using 12.3M.

Theo reports a 27% Gemini cache miss rate that raised costs roughly 50% above expectations and links a reproduction. The visible exchange with Phil Schmid does not establish which endpoint was tested.

Asana reportedly reduced its GPT-6.1 Sol browser-agent cost from $1.97 to $0.47 per run with an 89% cache hit rate. The report describes append-only prompt history and a Codex cache-measurement command.

Daniel Mac shared a script to compare Claude Code subagent cache lifetimes, then reverted to five minutes after finding no savings for his workload. He also corrected the hour-long setting from 60 to 1h.

SemiAnalysis treats subscription plans as model-specific credits and reports 4–5× more API-equivalent value for Anthropic plans than OpenAI’s. Task-cost-adjusted comparisons still favor Claude by 1.3–2.9×.

GPT-6.1 Sol scored 86.3% on MathArena BrokenArXiv, 78.6% on Braintrust problem solving, and third in Code Arena WebDev. A user also reports that it outperformed its mini-SWE result in Codex.

ARC Prize reports lower ARC-AGI-1 and ARC-AGI-2 scores for GPT-6 Sol than GPT-5.6 Sol at matched effort. Lower prices and fewer reasoning tokens cut per-task costs by over 70%.

Sonnet 5.5 ranks second on Artificial Analysis' Intelligence Index but uses about 193,000 output tokens per task at max effort. A comparison of AA Index results puts its cost at $7.60 per task.

Kevin Kern's single-task comparison favored Astra coordinating Opus 5.5 for quality. Solo runs were close, cost less and finished about three times faster; a rerun corrected for polling costs.

An independent DeepSWE test found Jev Router roughly matched GPT-6 Astra low on results but cost slightly more and took nearly five times as long. Practitioners also argue that request-level routing misses repository context and cache costs.

Experiments reported by Maxime Rivest found task-specific small classifiers outperforming frontier models on specialized decisions. One result says a 4-million-parameter BERT Tiny model beat Opus and Kimi after training on 10,000 examples.

OpenRouter launched Jev Router, which selects a model and reasoning effort per turn while weighing the cost of losing cached context. OpenRouter reports 237 of 423 tasks solved; invalid Jev outputs or timeouts fail without fallback.

OpenAI released GPT-6 Sol and GPT-6 Luna at API prices roughly half those of their GPT-5.6 predecessors. The models add controllable prompt-cache breakpoints and are rolling out in Codex and ChatGPT Work.

Teknium’s public evaluation says a Jev compaction strategy removes tool calls and eventually stops yielding savings. Repeated compaction can invalidate caches and increase total token costs, according to the critique.

Stagehand says adding Jev to its browser primitives cut median Act latency from 1.97 seconds to 0.46 seconds. Jev handles page-level choices and falls back to an LLM when uncertain.

Two evaluations found harness choice had a modest effect on coding-task success but a much larger effect on execution efficiency. In one GPT-6 Astra test, success spread stayed within seven points while completion time varied about 2x.

MiMo is reportedly livestreaming RL training for its V2.6 Pro and Flash models, publishing batch data, harness composition, reward curves, and infrastructure metrics. Reported cost figures list the trillion-parameter Pro run at about $493,000.

Polylane says it replaced role-specific sub-agents with one main agent and improved quality while reducing latency and cost. The report is a practitioner case study, not a general benchmark.

Fusion keeps a lead model in control while routing execution to a cheaper model in Devin CLI. Cognition reported 39% lower benchmark cost, and Artificial Analysis measured 43% lower cost with 31% faster runs at near-frontier scores.

Cognition says its SWE-2 coding model scored 50.0% on FrontierCode 1.1 Main, matching Fable 5.1 at 64% lower cost. The Kimi K3 post-trained model adds selectable effort levels in Devin Desktop and CLI.

Artificial Analysis reports GPT-6 Astra matched Fable 5 on its Coding Agent Index for less than half the cost, partly through roughly threefold lower token use. Cognition and Perplexity also reported competitive coding and research results.

Cohere released Parse 5, a document parser that returns machine-readable text, tables, forms, images, and bounding boxes. Cohere reports a 79.2 ParseBench score and prices it at $1.50 per 1,000 pages.

Glean says its runtime routes enterprise-agent work across more than 40 models using company context. It reports $0.58 per query and 78% user preference over Claude Cowork in a 180-person benchmark.

Together reports GLM-5.3 solved 87.6% of DeepSWE after four attempts for about $16, versus Fable 5 at 69.7% for $21.63. It estimates equal $100 budgets yield roughly 17 solved tasks for GLM-5.3 and three for Fable 5.

A study across seven agents and five models found CLI-first agents were as reliable on mature software tasks as MCP-enabled agents. The CLI setups cost 5–28 times less in the reported experiments.

OpenAI cut GPT-5.6 Sol API input and output rates from $5 and $30 to $4 and $20 per million tokens for three months. Subscription usage remains unchanged.

ARC Prize verified Gemini 3.7 Flash at 84.6% on ARC-AGI-2, at a reported $0.25 per task. Artificial Analysis’ AnalystAgent benchmark placed it at 60%, ahead of Claude Opus 5 and GPT-5.5.

OpenCode says it revised Go limits after DeepSeek raised prices. Its operator is testing hosting configurations intended to bring DeepSeek service closer to its prior price point.

Composio and Ante tests reported that the same models behaved very differently by harness. DeepSeek V4 Flash ranged from 47% to 67% task success and $0.019 to $0.104 per task across harnesses.

OpenRouter upgraded Auto Router to classify prompts into about 30 task types, then route by anonymized 7-day spend share and cost tier. OpenRouter says the max tier beat the old router across five benchmark domains.

A Microsoft Copilot trace analysis said 87% of LLM calls came from the agent, not direct user turns. Related posts warned token use and web requests can scale far faster than human prompt counts.

OpenCode’s thdxr said Go users spent $1.14 per day on DeepSeek V4 Flash last week. Wafer added a fast OpenRouter route, while Nous extended a 90% discount for the 0731 model.

Together says two V4 Flash attempts solved more DeepSWE tasks than one GPT-5.6 Luna attempt for roughly one-third the cost. Practitioners report Flash-0731 results vary sharply by harness and pass count.