Fresh context for agents. Rebuilt every 12 hours. Provenance by Synorb.
Updated 2026-11-17 00:00 UTCnext update 2026-11-17 12:00 UTC
Confirmed · 4 Manifests813 words

Your coding agent bill is now a line item. Price it per task

A benchmark puts median AI coding spend at $213 per engineer per week, and enterprise token data shows companies switching providers fast. The reason is structural: agentic work resends whole conversations and runs multi-agent graphs.

Key facts
  • The Larridin Benchmark puts median billed AI coding spend at $213 per engineer per week, with the 90th percentile at about $3,900 per month. Self-reported m:…0581
  • Databricks says enterprises using its Unity Gateway fell from 64% relying on a single model provider in July to 28% by November 2, while open-weight models' share of token traffic rose from 7.3% to 30.9%. Self-reported m:…5266
  • In a published Anthropic study cited by Aethir, multi-agent systems used about 15 times the tokens of a chat, and Gartner expects agentic inference costs per workflow to rise more than fivefold through 2028. Self-reported m:…4196
  • Together AI says long-horizon agent runs resend the entire conversation every step, so prefix caching is what saves most of that work. Self-reported m:…0012

The spend is real, and it does not buy output by itself

AI coding is a real line item now, and sooner or later finance will ask whether your number is high or low. That is how Larridin frames its benchmark, which puts the median billed AI coding spend at $213 per engineer per week, with the 90th percentile at about $3,900 per month. The benchmark's central finding is that more spend does not automatically mean more output: spend and output only correlate for AI-native engineers. At equal spend, deeply AI-native engineers ship about twice as much as partial adopters, and partial adopters see their marginal payoff drop by about half once spend passes roughly $600 per week. The deeply AI-native group has not yet hit a point where more spend becomes less productive. m:…0581

The benchmark helps answer whether your spend is high or low, Larridin says, but what it cannot do on its own is tell you whether the money is buying anything. That is the right question, because the cost structure of agentic work is different from chat. In a published Anthropic study cited by Aethir, multi-agent systems used about 15 times the tokens of a chat, and Gartner expects agentic inference costs per workflow to rise more than fivefold through 2028. Aethir also cites a Gartner forecast that up to 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025. The bill grows with the number of agents and the depth of their graphs, not just with the number of users. m:…0581 m:…4196

Why the tokens pile up

Together AI explains the mechanics. Every time an agent takes a step, the harness sends the model the entire conversation so far: the system prompt, every tool definition, the request, every file the agent has read, and every command it has run. That large input and small output pattern is characteristic of most agentic workloads. Prefix caching is what saves most of that work, but changing anything earlier, even a single character, means reprocessing everything after it. The easiest way to wreck the cache, Together AI says, is to put something in the system prompt that changes from turn to turn, like a current timestamp, the git status, or a working directory the agent keeps moving through. On a long run, that can be hundreds of thousands of tokens spent on changing one setting. m:…0012

The fix is partly about ordering: Together AI says the better approach is to put definitions at the end, where new content goes anyway, and to hand out cache memory in small blocks as each request needs them, the way an operating system handles RAM. If the next turn lands on a different node, that node has to process all earlier turns from scratch, and two nodes end up storing the same prefix. The company says harnesses like Claude Code, Codex and Pi can now connect with open models like DeepSeek-V4.1-Flash and GLM-5.3 while keeping existing workflows, which matters because the cheapest way to cut the bill is often to run the long, repetitive part of the work on a cheaper model. m:…0012

Enterprises are already switching

Databricks' new AI Index, built on aggregated, anonymised data from more than 10,000 companies using its Unity Gateway, shows enterprises are already acting on this. As recent as July, approximately 64% of enterprises using Unity Gateway relied on just one model provider; by November 2, that had fallen to 28% as more organisations adopted models from two or more providers. Open-weight models' share of enterprise token traffic grew from 7.3% in July to 30.9% in November, more than quadrupling in less than four months. In six weeks, Claude Opus 5.5 surged from 7.2% to 37.9% of proprietary-model tokens, lifting Anthropic's overall share from 57.6% to 68.8%. A new flagship proprietary or open-weight model has launched roughly every week since May 2026, or every 3.7 days when regional labs are included. m:…5266

The cost gap is large enough to justify the switching. In Databricks' coding benchmark, GPT-6 Luna and Claude Opus 5.5 both achieved an 89.1% task success rate, but Luna averaged just $0.05 per task compared with $1.62 for Opus 5.5, a roughly 32-times difference. Three open-weight models rank among the top 10 in the company's coding quality benchmark, including DeepSeek V4.1 and GLM 5.3, which both rank third with a 93.1% pass rate. The trap is that cheap tokens do not always mean cheap tasks: Gemini 3.8 Flash has a low token price, yet it averaged 140 turns per task and finished as one of the most expensive models in the benchmark at $11.97 per task. Aethir cites a UC Berkeley LLMCompiler result that treating calls as a dependency graph cuts latency by up to 3.7 times and cost by up to 6.7 times against a step-by-step ReAct agent. The decision for a founder is to budget per completed task, measure cost per task across providers, and fix the cache before buying more tokens. m:…5266 m:…4196

What is still unresolved

  • Whether the $213 median holds as more engineers become AI-native and spend rises further.
  • How much of the provider switching Databricks sees is driven by price versus capability, since the index reports shares, not reasons.

Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.

Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.

Run these Manifests yourself

Every Manifest cited here can be read in full with your own Synorb Key. Check the claims before you rely on them.

  1. Get a Synorb Key. Click Generate Free Agent Key at https://synorb.com/connect/start, or let your agent make a key: MCP tool synorb-get-demo-key (no header needed), or REST POST https://api.synorb.com/api/connect/demo-key with a JSON body such as {"client":"your-agent-name"}. The key is the api_key field of the answer. Save the key. Existing 72 Hour Agent Demo keys keep their issued limits. New keys require email OTP verification. An existing demo key lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; all of them are starter Beacons.
  2. Read each cited Manifest by id. REST: GET https://api.synorb.com/manifests/by-id/{manifest_id} with the header Authorization: Bearer <your Synorb Key>. Ids are strings; do not parse them as numbers. The reply holds the source, the Signal (claims) and the Brief.
  3. Check the claim. Match each sentence in the story to the claims in the Manifest it cites. The claim ids in the ledger point to them.
Cited Manifest ids (4)
  • 1794874263415630581
  • 1794864538050545266
  • 1794852294602424196
  • 1794866184200560012
for id in 1794874263415630581 1794864538050545266 1794852294602424196 1794866184200560012; do
  curl -s -H "Authorization: Bearer $SYNORB_KEY" "https://api.synorb.com/manifests/by-id/$id"
done

Provenance ledger

ManifestTitleSourceClaims usedPublished
1794874263415630581AI Coding Spend per Engineer: 2026 Benchmark Datalarridin-blog1794874546874632249, 1794874546881632254, 1794874546882632255, 1794874546894632260, 1794874546898632261, 1794874546876632251, 1794874546902632263, 1794874546884632257, 1794874546887632258, 179487454688963225911-16 00:00
1794864538050545266Introducing the Databricks AI Index: The state of enterprise AI tokenomicsdatabricks-blog1794864767612548613, 1794864767614548614, 1794864767609548610, 1794864767611548612, 1794864767605548608, 1794864767616548616, 1794864767615548615, 179486476761754861711-16 00:00
1794852294602424196What AI Agent Graphs Mean for GPU Compute Demandaethir-blog1794852843011426752, 1794852843013426753, 1794852843008426751, 1794852843014426754, 179485284302342676011-16 00:00
1794866184200560012Making open models work for long-horizon agentic workloadstogether-ai-blog-rss1794866360084561107, 1794866360088561109, 1794866360089561110, 1794866360090561111, 1794866360092561113, 1794866360093561114, 1794866360091561112, 1794866360099561119, 179486636009856111811-16 00:00

Related stories