Deepseek V4 Flash can reportedly process a one-million-token request for 14 cents, yet it also reportedly invents an answer 84 percent of the time when it does not know. Cheap context and trustworthy reasoning are different capabilities.
The price comparison is striking: $5 for GPT 5.6 Saul, $5 for Claude Opus 5, and $1.50 for Gemini 3.6 Flash. At 14 cents, longer agent workflows become much easier to justify economically.
Cheap context is not reliable reasoning
The reported efficiency gains are substantial. At one million tokens, cache use fell from 83.9 GB for V3 2 to 9.6 GB for V4.
V4 Flash uses a sparse architecture that combines selected compressed summaries, a dense 128-to-one skim of the full prompt, and standard attention over the latest 128 tokens. It can ingest a very large prompt without treating every part equally.
But ingestion is not dependable reasoning.
V4 Flash reportedly averaged 0.49 on a one-million-token multi-needle retrieval test, with reliable retrieval extending only to approximately 128,000 tokens. The cited Sparse Frontier study found that distributed, multihop work deteriorated at substantially lower sparsity than simple retrieval.
On Artificial Analysis’s Omniscience benchmark, V4 Flash scored minus 16 and reportedly invented an answer 84 percent of the time when it did not know.
Treat context capacity as an input budget, not a measure of trust.
Use the cheap model to narrow the problem
The practical architecture is selective. Let a cheap model read widely and extract the relevant slice. Then give that smaller problem to a stronger model for narrow reasoning and verification.
Neither tier is sufficient alone. The cheap model makes broad ingestion economical. The stronger model handles the work where accuracy matters. This design uses low-cost context without assuming the cheapest model can reason reliably across everything it can technically ingest.
Ramp shows what makes an agent useful
Ramp offers a production example built around measurable work.
In one reported run, Fable reduced CI time from a recalled 18-minute P50 to a 6-minute P50. It then returned over several days, inspected production data, and applied further improvements.
That is more useful than a one-shot coding demonstration. The agent worked against a measurable result, and production data could be used to check its work.
Ramp separates repetitive tasks from work whose steps are not known in advance. Fixed loops handle repetitive work. For dynamic workflows, Claude coordinates subagents and adversarial checks. The workflow changes with the uncertainty of the task.
Reliability comes from constraints and observability
Ramp’s agents receive limited credentials, including read-only access to BigQuery or Datadog. The team inspects individual traces to identify missing context, tools, or commands.
Ramp does not impose individual token budgets. It assigns cheaper models and batch or flex APIs to unattended work. Cost optimization follows the identification of work that is useful and verifiable.
Let cheap models read broadly, but only let agents act narrowly and require a result that another system, model, or engineer can verify.
Cheap inference moves the engineering effort into permissions, evidence, and failure inspection. Cheap context expands the possible workflow; constrained access, inspectable traces, measurable outcomes, and targeted verification make that workflow useful.