Per Job Cost

Per job cost is the total real money and time burned to ship one finished output from a frontier model. You start with the raw token math, layer on the failure rate that forces second and third calls, add the designer hours spent fixing drift or hallucinations, then factor in the context window inefficiency that inflates every prompt with endless background material. In 2026 production stacks this single number determines routing logic because a DeepSeek call at thirty cents per million can easily cost more than an Opus call at fifteen dollars if the cheap model needs three retries on a hard reasoning task.

Per job cost is not the price listed on any model card. Those numbers only show dollars per million input and output tokens. It is not even the same as per call cost because calls vary wildly by workload. A simple classification job on Claude 4.7 Haiku might finish in one shot for four cents while an agent run on the same model balloons to nine dollars after the model loses the thread at turn seventeen. The cheap token trap that teams fall into every quarter ignores this completely. They see DeepSeek V4 at thirty cents input and route everything there only to discover their per job cost quadrupled from rework.

Look at the research synthesis pipeline that ran at Figma throughout Q1 2026. The design systems team dropped ten quarterly reports plus competitor teardown decks into a single prompt then asked for a grounded insight summary with citations. Using Gemini 3 Pro at two fifty input and ten output the average job ate one point two million tokens for roughly four dollars and twenty cents. Retrieval stayed reliable so the summary needed only light designer review of eleven minutes at a fully loaded rate of ninety dollars per hour. Total per job cost landed at twenty dollars and seventy cents. When the same team tried routing the identical workload through DeepSeek V4 the token cost dropped to forty eight cents but the model missed key citations thirty seven percent of the time. Each failure triggered a full re prompt plus an extra thirty two minutes of designer cleanup. The per job cost jumped to thirty one dollars and twelve cents. Figma switched the router back to Gemini for any synthesis job over two hundred thousand tokens and never looked at per token pricing again.

Another concrete case hit Cursor users building agentic features in March 2026. A prompt as component task required strict JSON output, three tool calls, and perfect format compliance across a forty turn agent run. Claude 4.7 Opus at fifteen dollars input and seventy five output delivered a first try success rate of ninety four percent for an average per job cost of nine dollars and forty cents including minimal review. The same team routed identical jobs to Llama 5 405B hosted at one dollar sixty blended. Token cost fell to two dollars ten but success rate dropped to fifty one percent. The failures each demanded a full restart plus forty seven minutes of engineer debugging. Per job cost exploded to forty one dollars. Cursor updated their default routing to keep Opus for any prompt as component primitive and reserved Llama 5 for non critical classification only.

Calculate per job cost when you audit a production routing layer or before you lock vendor contracts for the next twelve months. Pull logs from the last ten thousand jobs, tag them by use case, measure actual token burn, multiply by current pricing, add observed retry rates, then attach the real hourly cost of whoever fixes the output. Run the spreadsheet every quarter because model behavior drifts, failure rates shift, and new cheaper models like whatever DeepSeek ships in July can change the math overnight. Do not calculate per job cost when you are personally sketching ideas inside Claude at two in the morning or when you are running one off image generation experiments where your own time is the only currency that matters. Those situations still run on gut feel and raw token price.

The teams that win in 2026 stopped shopping for the lowest price per token and started obsessing over the lowest price per finished job.