GPT-6 Luna Reads Slack for $0.007 a Thread: What That Number Hides

Seven-tenths of a cent. That is what Deepnote Research's Agent Initiative Benchmark recorded as GPT-6 Luna's mean cost to handle one workplace situation inside a simulated company - reading the messages, deciding whether anyone needed to be told anything, and replying in the right channel. Claude Fable 5.1, on the same fifty situations, with the same tools and the same 180-second response budget, averaged $3.02 for a score 1.1 points higher: 6.38 out of 10 against 7.50. The benchmark's own write-up calls the two models "430x apart in price," and the arithmetic holds: $3.02 divided by $0.0070 is 431.

GPT-6 Luna has been generally available since September 22, 2026, and the comparison went around the timeline on November 13 as a story about a cheap model catching an expensive one. It is a more interesting story than that, and a slightly less flattering one for the headline. Luna's own row in that table contains nine runs scored 10 out of 10 and nine scored zero, and the reason a proactive agent is hard to buy is not the average - it is the zeros.

What the benchmark actually ran

Agent Initiative is not another static question set. Deepnote built a fake software company called PageMantle - 71 employees, 100 enterprise customers, six years of connected records, about 10 GB across eleven services including Slack, Notion, Linear, HubSpot, Stripe, PostHog, BigQuery and Snowflake - and then dropped twelve frontier models into it as lurkers. Each model got the same instructions, the same tool access, the same company snapshot and the same time budget. The job was to read new Slack messages and decide whether to investigate, where to answer, and whether to answer at all.

The panel is small and deliberately awkward:

• Fifty situations across eight task types - Analyze (8), Verify (7), Coordinate, Explain, Diagnose, Decide and Guard (6 each) and Find (5).

• Forty of the fifty needed a message; ten were traps where the right move was silence.

• Two LLM judges, blind to which model produced the run, scoring 0 to 10 on intervention, evidence, tone and audience placement.

• Every model at high reasoning effort, with tool execution time excluded from the 180-second clock.

• Twelve models from six labs, 600 runs total, with 95% intervals from task-bootstrap resampling.

That design choice - forty help cases against ten silence cases - is why the table separates models in a way a chat benchmark never does. Most of what an always-on agent does all day is decide not to speak, and almost no public evaluation scores that.

Luna's row, in full

GPT-6 Luna finished seventh of twelve at 6.38. It sent a message in 38 of the 40 situations that required one, which is a better hit rate than Claude Sonnet 5.5's 37, and it did so in 50 seconds using 13.7 tool calls. Its cost per task, $0.0070, is the lowest in the field by a factor of nearly four against DeepSeek V4.1 Flash's $0.026 and by a factor of sixty against Claude Sonnet 5.5's $0.42.

The column that matters is the one Deepnote labels "Tens / zeros." GPT-6 Astra, the top scorer at 8.97, recorded 26 perfect runs and two zeros. Claude Opus 5.5 recorded 14 and two. Luna recorded nine and nine. In a benchmark where a zero means the agent either never spoke when it should have or spoke into a situation it had misread, one run in five for Luna was a write-off, against one in twenty-five for the two models above it. Averaged over fifty tasks, an agent that is excellent half the time and absent the other half looks like a B-minus. Deployed against a real team's Slack, it looks like a colleague who is either brilliant or silent, and nobody can predict which.

The 430x headline, with the arithmetic shown

The number is real and it is also the most favourable framing available, for two reasons worth stating plainly.

First, Claude Fable 5.1 is not A​nthropic's current model. Claude Fable 5.2 is, and it scored 8.17 on the same benchmark at $1.87 a task. Against the successor, GPT-6 Luna is 267 times cheaper and 1.79 points behind. That is a less quotable sentence and a more decision-relevant one, because nobody choosing a model this week is choosing Fable 5.1 unless they have a reason.

Second, the 1.1-point gap is a point estimate over fifty tasks. Deepnote is explicit that several models in the middle of its table sit inside overlapping intervals - it names Claude Sonnet 5.5, GPT-6 Luna, DeepSeek V4.1 Flash, Mistral Large 4 and Claude Haiku 5.5 as a cluster within 0.6 points of each other - and it plans a larger task set in a future version. The honest reading of the Luna-versus-Fable 5.1 pair is that the price gap is measured to three significant figures and the quality gap is measured to roughly one. Anyone quoting "only 1.1 points behind" as though the decimal were meaningful is quoting the interval, not the score.

The number nobody quotes: four out of ten

When silence was the correct answer, the median model in the field held back six times out of ten. GPT-6 Luna held back four. It intervened on six of the ten situations where the records already showed the matter had been handled - and, as Deepnote points out, those interventions were not gibberish. They were accurate, well placed, and unnecessary. A quality check that only reads the message will never catch that failure. Neither will a user satisfaction survey, because nobody complains about being told something true.

The benchmark puts a price on it anyway. Across the panel, speaking out of turn cost about 1.8 times as much as staying quiet when a message was needed. For a Slack agent whose entire premise is that it reads every thread, the failure mode is not a wrong answer that gets corrected. It is the slow erosion of the channel: once a team learns that a third of the bot's messages are noise, they stop reading any of them, and the value of the correct ones goes to zero with them.

There is a second, quieter failure in the write-up. In a data-access task, Luna found the record confirming that an access administrator had handed over to a successor, and then referred the employee to the former administrator anyway. Reading the right evidence and drawing the wrong conclusion is a different defect from missing evidence, and it is the one that survives a retrieval-quality fix.

Why seven-tenths of a cent changes the design, not just the bill

Cost per token is the wrong unit for this decision, because a proactive agent's cost scales with the number of threads it touches, not with the number of questions it is asked. Deepnote's own chart makes the point concrete: $7 per 1,000 tasks for GPT-6 Luna against $3,020 per 1,000 for Claude Fable 5.1. At 2,500 threads a month, the same figure the benchmark uses for Sol and Fable 5.2, Luna's bill is about $17.50.

That number is small enough to change what you build. At $3 a read, a proactive agent has to be selective before it starts - someone has to decide which channels are worth watching, which is exactly the judgment the model was supposed to make. At less than a cent a read, you can watch everything and let the model do the triage, which is the only version of the product that matches the pitch.

It also buys a second pass. Run GPT-6 Luna over all 2,500 threads for roughly $17.50, have it flag the fifty that look ambiguous, and send those to Claude Sonnet 5.5 at $0.42 each: $21 more, for a total near $39 a month. The all-Fable 5.1 version of the same coverage is $7,550. Whether that hybrid actually beats Luna alone is untested - the benchmark ran models, not harnesses, and Deepnote says so in its limitations - but the arithmetic is what makes the experiment affordable enough to run.

The plumbing arrived three weeks earlier

Cheap per-call reading is only half of what an always-on agent needs; the other half is a cheap way to ask a narrow question. On October 22, 2026, O​penAI took its Decisions API out of beta, and GPT-6 Luna is the only model it supports. The endpoint returns typed answers rather than prose - a probability that a condition is true, a choice from a fixed list, or a score against ordered levels - and O​penAI claims it is about ten times faster than asking the same model through the Responses API.

For a Slack watcher, that is the right shape. "Does this thread contradict the release date in the pinned message?" is a predicate, not a paragraph. The general-availability release raised the ceiling to 400 questions per request, added prompt caching for shared context, allowed images inside individual question instructions, and accepted structured data directly through an input_json content part. A single call can now evaluate an entire day of channel activity against a fixed evidence set instead of one message at a time.

Worth knowing before you design around it: the new caching applies to the shared evidence you send, not to Luna's rate card, and the model's own limits still apply - a 1,050,000-token context window, 922,000 maximum input, 128,000 output, and a May 18, 2026 knowledge cutoff. Function calling on Chat Completions only works with reasoning effort set to none; anything with built-in tools has to go through the Responses API.

Where the evidence stops

The benchmark is unusually candid about its own edges, and they matter for anyone about to put a model in front of a real team.

• Fifty tasks, with some categories holding only five to eight, so per-category rankings are tendencies rather than orderings.

• A fixed company snapshot whose teammates never reply, which measures one contribution rather than a month of accumulating context.

• One effort level. Every model ran at high; lower and higher settings were not tested, and Luna's default is medium, not high.

• Slack only. Email, documents and any action beyond sending a message are out of scope.

• Historical costs, not live quotes. The per-task figures are observations from the run, so a vendor price change moves them.

• No router was tested. Deepnote is explicit that its routing advice is a design recommendation extrapolated from model results.

There is also a tension between that table and the independent speed measurements. Deepnote clocked Luna at 50 seconds per situation at high effort. Artificial Analysis, running it at maximum effort, measured a time to first answer token of 79.33 seconds - an order of magnitude worse than the median for its price tier. Both numbers can be true, since they measure different things at different settings, but the pair is a warning: an always-on agent's perceived latency is set by the effort dial, and turning it up to chase the zeros costs seconds before it costs dollars.

Deploying a reader at this price

The practical version of this benchmark is a two-tier pipeline, and it is the reason the price gap matters more than the score gap. GPT-6 Luna at low or medium effort does the reading - every thread, every channel, typed decisions through the Decisions API. A threshold on those decisions escalates the ambiguous cases to a stronger model. The expensive model then handles tens of threads a month instead of thousands, and its cost stops being a budget line.

That pipeline is easier to run when both tiers sit behind one key. OrcaRouter exposes GPT-6 Luna, Claude Sonnet 5.5, Claude Fable 5.2 and more than 200 other models through a single OpenAI-compatible endpoint, at 0% markup - provider list price passed through, so a vendor price cut lands on our side the same day rather than waiting for a reseller to reprice. The routing DSL composes the tiers into one call (cheap first pass, escalate on a condition), automatic failover keeps the watcher alive if a provider degrades, and model fusion lets a panel answer together when a single classification is too consequential to trust to one model. None of that changes the numbers above; it just removes the integration work between them.

What would move this number

Two things would make this article obsolete: a GPT-6 Luna successor, or a second independent run of the same benchmark with a larger task set. The first is a matter of O​penAI's calendar. The second is already promised - Deepnote says a 50-task panel will be replaced by a larger sample in the next version, and its authors are clear that the current table should be read as a first measurement rather than a ranking.

Until then, the defensible conclusion is narrower than the timeline version and more useful. GPT-6 Luna is the first model cheap enough that a proactive agent can afford to be wrong about whether to look, and that is a real threshold - the difference between a product that samples its company and one that reads it. It is not yet reliable enough to be wrong about whether to speak. The four-out-of-ten silence score is the part of the table to design around: keep the read cheap, keep the write gated, and put the expensive model where the decision to interrupt actually gets made.