Almost every lab cut prices in the last four weeks. If you build or use coding agents, that sounds like the problem is solving itself. Reading the pricing pages closely, I think it's more interesting than that: per-token prices are dropping, but the pricing is increasingly shaped around how you send tokens: cache hits, prompt length, and which model does which step. A careless long session can still burn money on cheaper models.
Here's what changed, with sources, and what I take from it as someone building a VS Code coding agent.
What actually got cheaper
Claude Opus 5.5 (22 Sep) lists at $4 per million input tokens and $20 per million output, down from $5 / $25 for Opus 5 (Anthropic pricing). Its cache hits are priced at 0.05x base input, i.e. $0.20.
Claude Haiku 5.5 (7 Oct) is the big one: $0.10 / $0.50 per million for prompts up to 100,000 tokens, against $1 / $5 for Haiku 4.5 (announcement). Anthropic positions it for "summaries, compactions, database queries, and classification" and says it "pairs well with Opus 5.5 and Sonnet 5.5 as a subagent on coding work."
The same day, Sonnet 5.5 cache reads were halved from $0.20 to $0.10 per million, which Anthropic says makes Sonnet 5.5 "around 20% cheaper on most agentic work", because cache reads are a large share of what agents consume (announcement).
Gemini 4 Argon (13 Oct) launched at an introductory $2 / $10 per million, with cached input at 95% off and a 1M unified input-output token limit. After 30 November the price becomes $4 / $20 (Google).
The fine print is where agents live
Three details in those pages matter more to agent builders than the headline numbers.
1. The 100K cliff. Haiku 5.5 costs $0.10 / $0.50 only under 100,000 tokens. Above that it's $0.50 / $2.50 (model docs). Anthropic notes about 90% of requests to the previous Haiku fell under that line (announcement). A short classification call stays under it easily. A coding agent that has been running for an hour, resending its whole transcript and every old tool result, does not. The same model can be 5x more expensive per input token depending on how much history you drag along.
2. Cache pricing is now the main discount. Look at where the cuts landed: Sonnet's cache reads halved, Opus 5.5 cache hits at 0.05x, Argon's cached input at 95% off. The cheapest token is one you've already sent with an identical prefix. That rewards keeping the start of the prompt stable and penalizes anything that keeps rewriting it.
Claude Code's 2.1.299 release shows the tools adapting to this. A conversation over 200K tokens that sits idle for about 54 minutes is now compacted before its one-hour prompt cache expires, so your next message doesn't reprocess the whole history (Offsite's October digest). That's a feature built entirely around cache timing.
3. Cheap small models only help if something routes to them. Haiku 5.5 is explicitly sold as a subagent. But it scores 39.2% on Terminal-Bench 4.0 against 70.6% for Sonnet 5.5, as summarized by Vibe Coding Weekly #52, so it isn't a drop-in replacement. The same newsletter points to Michael Lynch's complaint that agents "use expensive models for gruntwork and never delegate to cheaper ones or escalate."
The best real-world cost data I found this month
Cockroach Labs published five months of running a coding-agent pipeline on its MOLT tools (21 April to 11 September): just over $135,000 in Claude tokens, around $84 per GitHub issue, more than a million lines of code landed with 7 reverts (Cockroach Labs).
What I found most useful is how they control cost. Planning now runs on their strongest model, the plan chooses the model for implementation, and work that struggles on Sonnet or Opus can escalate to a stronger one. They also run optimization rounds every couple of weeks after model switches.
One detail from coverage of the post stuck with me: one of their review agents loaded about 21,800 words of instructions, roughly 29,000 tokens, on every run before even reading the pull request, and a July audit flagged 23% of their skill files as removable without changing a gate (Reveneau's summary). Fixed per-run context is a cost line, even in a well-run pipeline.
What I take from this as an agent builder
- Track tokens per task, not price per token. A 20% list-price cut disappears if the session carries twice the history.
- Keep prompts short and prefixes stable. Tier cliffs like Haiku's 100K line and cache discounts both reward this. Compaction helps, but rewriting history also changes the prefix, so it's worth doing deliberately rather than constantly.
- Give the cheap model the cheap jobs. Summaries and compaction are exactly what Haiku-class models are pitched for. Planning and hard reasoning still deserve the strong model.
- Prices change on a calendar. Argon's price doubles after 30 November. Hard-coded cost estimates go stale, so keep per-million prices configurable.
That last point is personal: in CodeCook, the VS Code agent I'm building, the summary model is a separate setting from the agent model, and per-million prices are configurable for exactly these reasons. But the lesson applies to whatever tool you use. This month, the model you pick matters less than what you keep sending it.
Prices as listed on the linked pages on 21 October 2026. Check them before you rely on them; they move.
Top comments (2)
The 100K cliff is the detail most people will skip right past — the same model costs 5x per input token just because of dragged-along history. What worked for us was routing at the task level instead of the session level: the planner picks the model tier per job, so routine work never touches the strong model. Price-per-token is the marketing number; tokens-per-task is the actual bill.
"Price-per-token is the marketing number" is a better summary than anything in my post. Routing per task also has a nice side effect: a job that starts fresh on the cheap tier never inherits the long history, so it stays under cliffs like Haiku's 100K line almost by default. How do you handle escalation when the cheap tier gets stuck halfway through a job? Does the stronger model get the full trace of the failed attempt, or a fresh brief?