OpenAI’s GPT-6 family gives agent builders a 1.05-million-token context window, but the headline number conceals a consequential pricing boundary. The documentation for GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna says that prompts exceeding 272,000 input tokens are charged at twice the input and cache rates, while output is charged at 1.5 times the normal rate for the entire request.
That boundary changes how teams should think about long-context agents. A million-token window is valuable for exceptional jobs—large repositories, extensive legal records, lengthy research corpora, or long-running investigations—but it is not a license to carry every observation and tool result forever. Once a request crosses 272,000 tokens, one additional chunk can reprice all the tokens already in the prompt.
The operational lesson is simple: context length is now a metered resource with a pricing cliff. Agent platforms need context budgets, compaction checkpoints, and retrieval policies just as they need token limits and tool-call budgets. The quality question is no longer only “will this fit?” It is “which evidence deserves to remain in the expensive working set?”
The million-token headline has a smaller economic boundary
Astra, Sol, and Luna each advertise a 1,050,000-token context window, a maximum of 922,000 input tokens, and up to 128,000 output tokens. All three accept text and images and support the Responses API, structured outputs, function calling, prompt caching, file search, web search, and a broad set of agent tools. Technically, an orchestration system can feed any of them a very large working set.
The standard prices differ sharply. Astra lists input at $10 per million tokens, cached input at $1, and output at $50. Sol lists $2, $0.10, and $10. Luna lists $0.10, $0.01, and $0.50. Above 272,000 input tokens, each model’s documentation applies the long-prompt multipliers to the full request: 2x for input and cache, and 1.5x for output.
Consider an Astra request with 270,000 uncached input tokens and 10,000 output tokens. Ignoring tool fees, it costs roughly $3.20: $2.70 for input and $0.50 for output. At 273,000 input tokens, the same shape costs about $6.21: $5.46 for input and $0.75 for output. A 3,000-token addition nearly doubles the request cost because it moves the entire prompt into the higher tier.
The absolute numbers are lower on Sol and Luna, but the discontinuity remains. More importantly, agents rarely make one request. A workflow may plan, browse, inspect files, execute tools, recover from errors, and verify its answer over many turns. If the system repeatedly resends a bloated history, the premium appears again and again.
Why agents accumulate context so quickly
A chat interface usually has an obvious history: user and assistant messages. An agent’s working set is more complicated. It can include system instructions, tool schemas, retrieved documents, screenshots, repository maps, command output, structured state, intermediate plans, error traces, and summaries of earlier work. Each individual addition may look reasonable. Together they create context debt.
Tool results are a common source. A browser agent may retain full pages after extracting the relevant facts. A coding agent may keep entire files even after narrowing the change to a few functions. An operations agent may append repetitive logs on every retry. Large tool schemas and connector descriptions can also consume tokens before the task’s evidence is added.
Long-running agents face a second problem: relevance decays unevenly. An early constraint may remain critical, while a later tool result becomes obsolete after the agent changes direction. Keeping everything avoids the risk of deleting a needed detail, but it transfers the burden to the model and the budget. More tokens can also make important evidence harder to locate, especially when similar or contradictory observations remain in the prompt.
A large window therefore solves capacity, not curation. It lets a system process unusually large evidence sets when the task demands it. It does not decide what should be retained, summarized, retrieved again, or discarded.
Build a context budget before the prompt
The 272,000-token threshold should become an explicit budget line in agent orchestration. Teams do not need to target the exact boundary; token counts vary with formatting, images, tool definitions, and serialization. A safer policy reserves headroom for the next action and for output.
For example, an agent could set a soft compaction trigger at 220,000 input tokens and a hard approval threshold at 260,000. At the soft trigger, it summarizes completed work, removes superseded tool output, and replaces large documents with references plus extracted evidence. At the hard threshold, it either splits the job, retrieves only the necessary sections, or deliberately accepts the long-context premium because preserving the full corpus is worth the cost.
A practical context budget should separate several classes of content:
- Immutable instructions: system policy, task requirements, permissions, and output contracts that must survive every compaction.
- Active state: the current plan, unresolved questions, tool handles, and facts needed for the next few actions.
- Evidence: quotations, code fragments, records, and observations that support the result and may need citations.
- Recoverable material: documents, files, and prior tool output that can be fetched again by identifier instead of remaining inline.
- Disposable traces: superseded plans, duplicate pages, successful low-level command output, and errors that no longer affect the path forward.
This classification gives compaction a policy. Immutable instructions are preserved verbatim. Active state is kept concise. Evidence is reduced carefully, with provenance. Recoverable material becomes pointers. Disposable traces are removed.
Compaction must preserve proof, not prose
Naive summarization can lower token counts while damaging an agent’s reliability. A free-form summary may omit a negative finding, blur which source made a claim, or convert uncertainty into certainty. For high-stakes workflows, context compaction should produce structured state rather than a narrative recap.
A useful checkpoint records the objective, constraints, completed actions, verified facts with source identifiers, unresolved conflicts, failed approaches, pending decisions, and artifacts that can be reopened. Exact strings—IDs, error messages, code symbols, contractual language, and numeric limits—should remain verbatim when correctness depends on them. The system can summarize surrounding explanation without summarizing away the evidence.
Compaction should also be testable. After creating a checkpoint, ask a validator whether every active constraint and unresolved question is represented. For coding agents, verify that referenced files and symbols still exist. For research agents, ensure every conclusion retains a traceable source. If the compacted state cannot support audit or recovery, it is too aggressive.
Retrieval beats permanent residency
The GPT-6 models support file search and web search, which makes selective retrieval a natural alternative to permanent prompt residency. Instead of attaching an entire handbook to every turn, store it externally and retrieve the relevant passages for the current decision. Instead of preserving all command output, retain the command, exit state, key findings, and a handle to the full log.
Retrieval has costs of its own: indexing, latency, tool fees, and the risk of missing relevant material. That is why the best design is hybrid. Keep small, high-value constraints in the active context; store the bulk corpus behind searchable references; and promote retrieved evidence into the working set only while it remains relevant.
Prompt caching helps with stable prefixes. Astra’s cached-input rate is one-tenth its uncached rate, Sol’s is one-twentieth, and Luna’s is one-tenth. But the long-prompt rule still doubles cache rates above the threshold. Caching reduces the cost of repeated stable material; it does not justify unlimited histories, nor does it solve attention and relevance problems.
Split work when the corpus is naturally separable
Some tasks genuinely need more than 272,000 tokens of source material. The right response is not always compression. If the corpus separates cleanly, staged analysis can be more reliable: analyze components independently, produce evidence-bearing summaries, then synthesize those summaries in a smaller final context.
A repository review can divide by service or package. A contract review can process exhibits independently before comparing obligations. A research agent can group sources by claim and merge the resulting evidence tables. The split must respect cross-document dependencies; otherwise, local summaries may miss contradictions that only appear globally.
When relationships across the full corpus are the task—tracing a theme through hundreds of documents, for example—the premium may be justified. The point of a budget is not to prohibit long context. It is to make the choice visible, measured, and attached to business value.
Measure context efficiency at the task level
Agent evaluation should record peak input size, the number of requests above the threshold, cache-hit volume, compaction count, retrieval calls, and total successful-task cost. Those metrics reveal whether a workflow uses the large window for exceptional evidence or simply leaks state.
Quality must be measured beside savings. Compare an uncompressed route with a budgeted route on factual accuracy, constraint retention, tool success, latency, and human acceptance. A cheaper prompt that forgets a requirement and forces a rerun is not efficient. A larger prompt that preserves crucial cross-document evidence may be the right design.
OpenAI’s million-token GPT-6 window expands what a single agent request can inspect. The 272,000-token price boundary clarifies the engineering tradeoff: capacity is abundant, but active context is not free. Teams that treat context as governed state—budgeted, compacted, retrievable, and auditable—will capture the benefit of long windows without paying the premium on every step.
