September 2026 Frontier Routing: Astra, Fable, 3.8 Flash, Grok 4.6.

I do not pick one frontier model in September 2026. I pick by job: GPT-6 Astra for hard coding and attended computer use, Claude Opus 5 first on Claude routes with Fable 5.1 only after Opus at high effort still fails, Gemini 3.8 Flash for cheap volume and Antigravity, Grok 4.6 under 200k for Grok-native and Cursor subagents. Always-on ops is Grok Bot — xAI's teammate product, not a model ID — I run one primary that knows the business, then additive specialists.

I'm William Spurlock — founder, AI Systems Architect, and Fractional AI CTO. I've built 600+ automations with 500+ still live, spent 20,000+ hours on agentic systems, and helped clients delete 35,000+ hours of busywork. I pay token bills. I do not collect a weekly religion.

If you want last spring's three-vendor map, I already wrote the May 2026 frontier comparison. That post is history. This one is the board I am running on September 5, after Fable 5.1 on the 1st, 3.8 Flash on the 2nd, Astra on the 3rd, and Grok 4.6 sitting since August 12.

The names so I do not flatten them:

Vendor	Use this	Role	Do not flatten into
OpenAI	GPT-6 Astra (gpt-6-astra)	Flagship, September 3	GPT-5.6 Sol / Terra / Luna stay the cheaper stack
Anthropic	Claude Fable 5.1 (claude-fable-5-1)	GA Mythos-class, September 1	Not a replacement for Opus
Anthropic	Claude Mythos 5.1 (claude-mythos-5-1)	Same weights, invite-only	Not a public default
Anthropic	Claude Opus 5 / Sonnet 5 / Haiku 4.5	Default complex / volume / cheap	Opus 4.8 is a Fable fallback, not the flagship
Google	Gemini 3.8 Flash (gemini-3.8-flash)	Current Flash, September 2	3.7 Flash = efficiency fallback; 3.1 Pro = preview
xAI	Grok 4.6 (grok-4.6)	Flagship model, August 12	Not Grok Bot
xAI	Grok Bot	Always-on agent product, August 11	Not Cursor cursor-grok-4.6-xhigh-fast

I will not write "the new GPT" and leave it there. Astra is one seat. Sol still exists. Fable is a promote, not an Opus swap. 3.8 Flash is a Flash swap, not a flagship. Grok 4.6 is a model ID. Those five sentences are the week.

---

Which frontier model should an operator use in September 2026? #

None of them as a default for every job. The operator pick is a board: coding, computer use, and cheap subagents get model IDs; always-on ops gets a product seat. If someone asks me "which model" and expects one string, they are shopping a keynote. I am shopping an invoice.

The board I am taping next to the router today:

Job	First pick	ID I pin	Promote / stay / refuse
Hard OpenAI coding, long Codex, messy repo	GPT-6 Astra at high	gpt-6-astra	xhigh / max only if I am reading the session
Claude coding that already passes	Claude Opus 5	claude-opus-5	Do not hop
Claude coding Opus fails at high effort	Claude Fable 5.1	claude-fable-5-1	Opus 4.8 if I need an older Claude, not a newer myth
Google / Antigravity coding	Gemini 3.8 Flash	gemini-3.8-flash	Keep 3.7 on cheap loops
Attended computer use, OpenAI lane	GPT-6 Astra	gpt-6-astra	Human on the desktop. Extra tool fees.
Attended computer use, Google lane	Gemini 3.8 Flash (Preview)	gemini-3.8-flash	Sandbox. Not a standing OS write path.
Cheap Cursor / n8n subagents	Gemini 3.8 Flash or Grok 4.6 under 200k	gemini-3.8-flash / grok-4.6	Not Astra. Not Fable.
Claude cheap glue	Claude Haiku 4.5	claude-haiku-4-5	Sonnet 5 if the rubric needs a workhorse
OpenAI volume that already passes	GPT-5.6 Sol / Terra / Luna	Existing 5.6 IDs	Do not "upgrade" a labeler
Invite-only / Fairwind-only SKUs	Nowhere	—	Mythos 5.1 and 3.8 Flash Cyber are not on this machine
Always-on ops	Grok Bot (product)	Not a model ID	See the board note above — not a Cursor pin

That table is the post. The rest is receipts so nobody "upgrades" a classifier into a $50-output heater, or files a ticket asking me to turn on Mythos.

I already wrote the day-one cards this week. Use them when you need the spec, not a second religion:

• GPT-6 Astra operator spec card
• Where Astra goes and where it doesn't
• Astra field guide for this week
• Don't swap Opus 5 for Fable 5.1
• Gemini 3.8 Flash operator swap
• Grok 4.6 and the 200k cliff
• Grok 4.6 Extra High Fast in Cursor

This page is the picker. Those pages are the cards.

I keep vendor lanes. I do not move a Claude route to Gemini because a Terminal-Bench row moved. I do not move a Flash loop to Fable because Fable is new. Cross-vendor "upgrades" are how you break evals you already trust.

---

What do I pick for coding? #

Astra on the hard OpenAI / Codex jobs. Opus 5 first on Claude. Fable 5.1 only after Opus at high effort still fails. 3.8 Flash on Antigravity and Google coding agents. Grok 4.6 when the caller is already Grok. That is four IDs and one promote rule. It is not "use the newest name."

I already run a daily split between Cursor, Claude Code, and Antigravity. The coding assistant showdown and the Cursor / Claude Code daily workflow are still the tool-picker posts. This week did not ship a new IDE. It shipped model IDs those tools can call.

How I seat a coding job on September 5:

Coding job	First ID	Effort I start at	Why that seat
Long Codex refactor, notes across windows, messy repo	gpt-6-astra	high	API page built Astra for hard end-to-end coding. Launch post puts the updated Codex harness 1.9x faster on Mind2Web than the current Sol setup.
OpenAI volume that already passes	gpt-5.6-sol / Terra / Luna	Existing	There is no GPT-6 cheap trio. Sol stays.
Claude hard reasoning that already passes	claude-opus-5	high	Fable 5.1 overview says start on Opus 5.
Claude job Opus fails at high	claude-fable-5-1	high (API default)	Promote, not a default swap. Same $10 / $50 sticker as Astra. Cache reads are $0.25.
Antigravity / Google-native coding agent	gemini-3.8-flash	MEDIUM or HIGH	Current Flash. Default on antigravity-preview-05-2026. Intro $0.75 / $3.75 through December 31, 2026.
Google classifier already good enough	gemini-3.7-flash	LOW	Official efficiency fallback. 3.8 works harder and burns more tokens.
Grok-native coding in a caller I own	grok-4.6	high, xhigh only if I sit with it	500k window. $2 / $6 under 200k. No official grok-4.6-fast.
Cursor disposable subagent	cursor-grok-4.6-xhigh-fast or gemini-3.8-flash	Task slug / Flash enum	A Task dies when the job dies. It does not know the shop.

Official coding receipts I will cite, not invent, from OpenAI's September 3 comparison tables:

Eval (OpenAI, Sep 3)	GPT-6 Astra	GPT-5.6 Sol	Claude Fable 5.1	Gemini 3.8 Flash
Terminal-Bench 4.0	57.9%	37.3%	55.8%	19.1%
DeepSWE v1.1	74.1%	72.7%	67.4%	73.8%
BenchCAD	95.9%	83.3%	84.3%	—
Artificial Analysis Intelligence Index v4.1.1	61.2	60.9	65.7	58.7

Read that like an operator. Coding is not a wipeout. DeepSWE is a photo finish. Fable 5.1 still wins the intelligence index in OpenAI's own chart. Astra's gap that earns a new ID is computer use and long science loops, not "delete every other coding model."

Three coding rules I will not break this week:

1. Do not make Astra the default on every OpenAI node. Sol, Terra, and Luna still exist. A labeler does not need $50 output.
2. Do not make Fable 5.1 the default in Cursor. I wrote that rule the morning after the September 1 ship. A switch hides the jobs Opus already passed.
3. Do not point a Flash ID at an Astra job because the windows look similar. 1,048,576 versus 1,050,000 is not the difference. Output cap is: Astra 128,000, Flash 65,536.

If the job is "write the brand site," Astra can draft and QA. ChatGPT Sites is a beta host, not my production stack. I already said that on the Astra spec card. For the week-of prompts and fences, use the Astra field guide.

How I decide a coding hop on a live job:

Signal	Stay	Hop
Opus 5 at high passes the eval	Stay on Opus 5	—
Opus 5 at high fails twice on the same rubric	—	claude-fable-5-1
Sol already ships the patch	Stay on 5.6	—
Sol burns three retries on a messy repo	—	gpt-6-astra at high
Antigravity default is doing the work	Stay on 3.8 Flash	Do not cross-wire Astra into Antigravity
Cursor Task is a file edit	cursor-grok-4.6-xhigh-fast or 3.8 Flash	Not a flagship sticker

I log the hop. If I cannot write one sentence for why I left the first ID, I did not earn the second ID. That log is how I reverse a promote next week without guessing.

---

What do I pick for computer use? #

Astra on the OpenAI lane, attended. 3.8 Flash Preview on the Google lane, sandboxed. I do not have a standing unattended desktop writer, and I will not write one from a launch post. Computer use is a tool with its own fees, its own stop conditions, and a human on send.

What I will actually pin:

Lane	Model	Status I treat it as	Operator fence
OpenAI	gpt-6-astra	Generally deployed tool on Responses	Human watches the desktop. Search and computer-use calls bill extra on top of $10 / $50.
Google	gemini-3.8-flash	Computer use = Preview on the Gemini API page	Sandbox. Not a standing OS write path.
Anthropic	Opus 5 / Fable 5.1	Not a computer-use card I am publishing this week	I will not invent a Fable clicker.
xAI	grok-4.6	Model ID	Pin the ID. Do not invent a Fast SKU.
Enterprise ChatGPT	Astra	Off by default	Admin enablement. Early Model Access does not carry over.

The Astra launch post is the receipt that moved computer use for me: OSWorld 2.0 offline partial at 72.6% in about 40 minutes per task versus Sol at 65.7% in about 75 minutes. That is the gap I will pay $10 / $50 for. I will not flatten it into "Astra is cheaper." Price per task can fall if retries die. Price per million does not.

Computer-use rules I run this week:

1. Attended or off. A loop that clicks a real desktop without me in the room is a different product than a coding agent.
2. Responses, not a copied Chat Completions payload, on the OpenAI lane. The API page lists computer use on Responses. If the tool is not on the endpoint, the pin is theater.
3. Preview means preview on Gemini. I already wrote that in the 3.8 Flash swap. Preview is not a standing write path onto a client's machine.
4. No exploit write-ups. Astra is the first broadly deployed OpenAI model to hit the Critical cybersecurity threshold on the Deployment Safety Hub. Daybreak sits first. I will not paste, "summarize," or reconstruct an exploit path from a system card.
5. Count the tool fees. Token stickers are not the whole bill when the model is clicking.

What I will not do, even if a Slack says "just let it drive":

• I will not turn computer use on for a classifier.
• I will not publish a ChatGPT Site to the open web from an Enterprise workspace as a computer-use demo.
• I will not treat Daybreak access as something I already have.
• I will not route anything that smells like exploit development. Nowhere. I will not route it.

If the job is "the agent should keep working when I close the lid," that is not a computer-use model pick. That is a product seat, and I already named it once above.

A computer-use pass I will accept this week has four checks. Miss one and the run is a fail, even if the screenshot looks clever:

1. I started it. No overnight clicker on a real desktop.
2. I can stop it. A hung OSWorld-style loop is a failed run, not a "long think."
3. I can see the bill. Token sticker plus tool fees, written down before the session.
4. Nothing left the sandbox I did not name. Preview stays preview. Enterprise stays off until an admin flips it.

I will test Astra computer use on a disposable VM with me in the chair. I will not claim a client desktop win from OpenAI's 40-minute OSWorld number. That number is their bench. My pass is the four checks.

---

What do I pick for cheap subagents? #

Gemini 3.8 Flash on Google volume, Gemini 3.7 Flash on the loops that are already good enough, Grok 4.6 under 200k when the caller is Grok or a Cursor Task, Haiku 4.5 on Claude glue, Sol / Terra / Luna on OpenAI volume. Astra and Fable stay off this row. Cheap is a seat. Flagship stickers are not a personality.

The cheap board:

Cheap job	ID	Sticker I budget	Cliff I watch
Google agent / Antigravity default / cheap-but-strong coding loop	gemini-3.8-flash	Intro $0.75 / $3.75 per 1M through December 31, 2026, then $1.50 / $7.50	Thinking tokens. 3.8 works harder than 3.7. Output cap 65,536.
Google high-QPS labeler, draft, classify	gemini-3.7-flash	Same intro family	Do not "upgrade" a 10k-call/day loop to 3.8 without a token log.
Cursor disposable Task	cursor-grok-4.6-xhigh-fast	Cursor plan, not the xAI API card	A Task dies when the job dies. It does not know the shop.
Grok API / n8n Grok node, prompt under 200k	grok-4.6	$2 / $6 per 1M; cached $0.50	Whole request doubles at 200k. Confirm on the live sheet.
Grok prompt at or over 200k	Still grok-4.6, or split the prompt	$4 / $12 class if the double holds	Do not dump a repo "just in case."
Claude glue, subject lines, high-volume tags	claude-haiku-4-5	Cheap Claude seat	Do not put Fable on subject lines.
Claude volume drafts	claude-sonnet-5	Workhorse	Stay here if the eval passes.
OpenAI standing volume	GPT-5.6 Sol / Terra / Luna	Cheaper than Astra	Do not hop to gpt-6-astra because the flagship is new.

Google's own warning, in the September 2 launch post and the Gemini API model page, is that 3.8 Flash takes extra reasoning steps and extra tool calls on complex tasks. That is a bill. I set thinking_level to LOW on classifiers and MEDIUM / HIGH on agents. I deleted thinking_budget. An integer budget on a 3.8 payload is a copied 3.7 mistake.

Grok 4.6 cheap math I actually run, from the Grok 4.6 docs and the August 12 announcement:

Prompt size	What I assume	What I refuse
Under 200k	$2 in / $6 out per million, cache $0.50 when it hits	xhigh on a spam tag
At or over 200k	The whole request doubles	A 400k "just in case" dump
Invented grok-4.6-fast	No official ID	I send the node back

Astra cheap math I will not pretend exists: Standard is $10 / $50, cache $1, cache writes $12.50. Cross 272K input and the full request doubles input/cache and multiplies output by 1.5. Fast is 2x. Batch and Flex are half. That is the API model page. A subagent that classifies tickets does not belong on that card.

Fable cheap math is the same trap with a prettier cache: $10 / $50 sticker, cache reads at $0.25. Cache-heavy Claude loops can beat Astra on the invoice. They still should not classify spam. Haiku 4.5 exists for a reason.

If you want the habit behind the cheap row — name the hours, name the token path, name the failure cost — I already wrote how to calculate the ROI of AI automation before you build. This board is the September inputs. The Grok cliff write-up is pin the ID, watch 200k. The Astra surcharge write-up is where Astra goes.

Cheap-subagent pass/fail I actually use:

Check	Pass	Fail
ID matches the cheap seat	Flash, 3.7, Grok under 200k, Haiku, 5.6	Astra or Fable on a labeler
Effort matches the job	LOW / low / no xhigh	HIGH or max "just in case"
Tokens per successful task logged	I can name last week's number	"It felt cheap"
Retry count	One pass, or a named second pass	Silent three-retry loops
Caller	n8n / MCP / Cursor Task I own	A copied payload with last month's ID

I will not claim a dollar-per-draft winner from a vibe bake-off. I will claim the ID, the cliff, and whether the eval passed. If you want a same-brief bake-off later, bring a rubric. Do not bring a winner you invented on the flight home.

---

How do prices and windows change the pick? #

The sticker is not the job. The cliff is. Astra has a 272K surcharge. Grok 4.6 has a 200k whole-request double. 3.8 Flash has an intro clock and a 65,536 output cap. Fable's real number is the $0.25 cache read. I pick the ID, then I pick the cliff I am willing to stand on.

The window and price card I am running today:

Seat	ID	Window / max out	Standard sticker	Cliff
OpenAI flagship	gpt-6-astra	1,050,000 / 128,000	$10 / $50; cache $1; cache write $12.50	>272K input: 2x in/cache, 1.5x out on the full request. Fast = 2x.
OpenAI cheap stack	GPT-5.6 Sol / Terra / Luna	Existing 5.6 windows	Cheaper than Astra	Do not flatten into GPT-6
Anthropic promote	claude-fable-5-1	1M / 128K	$10 / $50; cache read $0.25	Always-on thinking. 2x input vs Opus 5 ($5 / $25).
Anthropic default complex	claude-opus-5	1M / 128K	$5 / $25	First pass. Stay if the eval passes.
Anthropic volume / cheap	claude-sonnet-5 / claude-haiku-4-5	Existing	Workhorse / glue	Do not promote these to Fable for vibes
Google current Flash	gemini-3.8-flash	1,048,576 / 65,536	Intro $0.75 / $3.75 through Dec 31, 2026	Then $1.50 / $7.50. thinking_level LOW / MEDIUM / HIGH. minimal errors.
Google efficiency	gemini-3.7-flash	Same Flash window	Same intro family	Keep on high-QPS loops
Google preview reasoner	gemini-3.1-pro	Preview	Preview card	Not this week's volume ID
xAI flagship	grok-4.6	500,000	$2 / $6 under 200k; cache $0.50	Whole request doubles at 200k. Effort adds xhigh. No Fast ID.

Worked Astra example I will put on the finance sheet. Assume Standard, no cache, no Fast, no tool fees, from the same API page:

Request	Input	Output	Math	Ballpark
Under the line	200,000	8,000	0.2 × $10 + 0.008 × $50	$2.40
Over the line	400,000	8,000	Full request at 2x in / 1.5x out: 0.4 × $20 + 0.008 × $75	$8.60
Same 400K, max bloats out to 40,000	400,000	40,000	0.4 × $20 + 0.04 × $75	$11.00

The jump is not "twice the tokens, twice the money." Crossing 272K reprices the tokens you already sent. If a Cursor or n8n route dumps a repo "just in case," you pay the surcharge even when the model only needed the failing file.

Worked Grok 4.6 example I run the same way. Assume the docs card and the 200k double:

Request	Input	What I budget
80k prompt, 4k out	Under the line	0.08 × $2 + 0.004 × $6 = $0.18
220k prompt, 4k out	Over the line	Whole request at the doubled card if the live sheet still says so
Same 220k with xhigh	Over the line + thinking	The cheap seat just stopped being cheap

I do not write "Grok" next to a raw 500k. I write the ID, then the integer, then the caller. I do not write "Gemini" next to a raw million. I write gemini-3.8-flash, then 1,048,576 in, then 65,536 out.

Knowledge cutoff on Astra is April 30, 2026 on the API page. Today is September 5. Fable 5.1, 3.8 Flash, and Astra's own launch are after that cutoff unless I attach docs or turn on web search. I will not ask Astra to "recall" this board from training. I will paste the board.

---

Which effort knobs do I actually set? #

Medium or high on flagship work I will read. Low on classifiers. xhigh and max are rare, attended, and named. Effort is how a $2 card still hurts and how a $10 card becomes a heater.

Model	Knob	Values I will set	Default I start at	What I refuse
GPT-6 Astra	reasoning.effort	low, medium, high, xhigh, max	medium or high	max on a classifier. Fast mode studio-wide.
Claude Fable 5.1	Effort / thinking	Always-on thinking. Claude Code defaults High. claude.ai defaults Medium. API default high.	API high on a promote	Making Fable the Cursor default
Claude Opus 5	Effort	High on the first hard pass	high	Promoting before the high-effort fail
Gemini 3.8 Flash	thinking_level	LOW / MEDIUM / HIGH	LOW cheap, MEDIUM/HIGH agents	thinking_budget integers. minimal.
Grok 4.6	Effort	low, medium, high, xhigh	high on messy judgment	xhigh on every Task "just in case"

OpenAI's evals in the Astra launch post are "maximum at any effort." That is how you print a chart. That is not how I price a standing agent.

Grok xhigh is 4.6-only. There is no official Fast ID next to it. If a contractor writes grok-4.6-fast into an n8n node, I send it back. Invented IDs are how weekend invoices happen.

Gemini thinking_level replaced thinking_budget. I already wrote the swap. If your payload still sends an integer budget, you are a string edit behind the current Flash — not a research project.

---

What will I not flatten this week? #

I will not flatten a flagship, a promote, a Flash, a cheap stack, an invite-only twin, and a teammate product into "the new model." Launch week is when operators lose IDs. I keep them.

The flatten list I am refusing:

Flatten	What it actually is	What I do
"Just use Astra"	OpenAI flagship at $10 / $50 with a 272K surcharge	Add a lane. Keep Sol.
"Switch to Fable"	GA Mythos-class promote at 2x Opus input	Promote after Opus at high fails.
"Mythos 5.1" as a public default	Invite-only twin. Same weights, different safeguards. Project Glasswing.	I do not have it. I do not route to it. Distinct from the April 2026 Mythos post.
"3.8 everywhere"	Current Flash that works harder and burns more tokens	Swap coding/agent IDs. Keep 3.7 on cheap loops.
"3.8 Flash Cyber"	Fairwind-only defender SKU	We do not have Fairwind. Public ID stays gemini-3.8-flash.
"Grok" as one noun	Model grok-4.6 plus a separate product	Pin the model. Do not call the product a model ID.
cursor-grok-4.6-xhigh-fast as a teammate	Disposable Cursor Task	Cheap subagent seat. It forgets the shop.
"GPT-6 Sol"	Does not exist	Cheap OpenAI stack is still 5.6.
Gemini 3.1 Pro as the Flash	Preview reasoner	Preview stays preview.
Opus 4.8 as the secret flagship	Fable fallback	Older Claude when I need one, not a new default.

I do not have Claude Mythos 5.1. I do not have Gemini 3.8 Flash Cyber. I will not write as if I do. Anthropic said Fable and Mythos are the same weights with different safeguards on the September 1 announcement. Google said Cyber is a Fairwind SKU on the 3.8 launch post. Access is the operator fact. A blog post is not an invite.

I also will not flatten this board into an agentic operating system. If you want the OS layer — standing jobs, shared picture, approvals — I already wrote what an agentic OS means day to day and what agentic AI is in 2026. This page is which brain I call. The OS is which job stays standing.

---

How does this update the May 2026 frontier pillar? #

The method stays. The names do not. The May 2026 Anthropic / OpenAI / Google pillar still wins as a way to compare context, price, and the job. The IDs in that piece are dated. Use this board.

What changed since May that I will actually act on:

May habit	September 5 fact	Operator move
Compare three vendors	Four vendors, plus a product seat that is not a model	Add xAI as a model lane. Do not add it as a fourth "chat tab."
Flagship vs cheap stack	Astra is the OpenAI flagship. 5.6 is still the cheap stack.	Do not delete Sol.
Claude default	Opus 5 is still the default complex. Fable 5.1 is the promote.	Do not swap Cursor defaults.
Google Flash	3.8 Flash is current. 3.7 is the efficiency fallback. 3.1 Pro is preview.	Swap the Flash ID. Keep the cheap loop.
"One best model"	The week shipped four names people will mash together	Pin the board.

Antigravity is still Google's multi-agent IDE. The Antigravity agents blueprint and the Antigravity 2 subagent recipes are still the maps for that surface. Astra is not an Antigravity default. 3.8 Flash is. Do not cross-wire the IDs because both launched this week.

I am not declaring an era. I am declaring a pin table.

---

What does a Tuesday look like on this board? #

Five jobs, five IDs, one human on send. I am not running a model bake-off at lunch. I am running the shop.

A Tuesday I will actually recognize:

Time	Job	ID I call	What I refuse
8:10	Overnight classifier backlog	gemini-3.7-flash at LOW, or Haiku 4.5	Astra. Fable. xhigh.
9:00	Client repo that failed Sol twice	gpt-6-astra at high in Cursor / Codex	Fast mode. max. A 400k dump.
11:00	Claude Code pass that already works	claude-opus-5	"Just switch the default to Fable."
2:00	Same Claude job, eval still red at high	claude-fable-5-1, logged	Making Fable the Cursor default
3:30	Antigravity agent on a Google route	gemini-3.8-flash at MEDIUM	Wiring Astra into Antigravity because both shipped this week
4:15	Disposable Cursor Task, file-level	cursor-grok-4.6-xhigh-fast	Treating the Task as a coworker that remembers the shop
5:00	Attended computer-use probe	gpt-6-astra on Responses, me in the chair	Unattended desktop. Exploit-shaped asks.

n8n pins I will write down, not "upgrade in place":

Node job	Model field	Notes I leave on the node
OpenAI hard agent	gpt-6-astra	Responses. effort high. No Fast. Watch 272K.
OpenAI volume	gpt-5.6-sol (or Terra / Luna)	Do not rename to Astra on a Friday.
Claude first pass	claude-opus-5	Promote only after a named fail.
Claude promote	claude-fable-5-1	Cache read $0.25 is the cost story.
Google agent	gemini-3.8-flash	thinking_level, not thinking_budget.
Google cheap loop	gemini-3.7-flash	Keep. Measure tokens/task for a week.
Grok caller	grok-4.6	Stay under 200k unless I approve. No Fast ID.

MCP is a tool pipe, not a model. If the server already works, I change the model string on the caller. I do not rebuild the server because a flagship shipped. If the tool is computer use, the caller is Responses on Astra or Preview on 3.8 Flash. A copied Chat Completions JSON that "should work" is how you ship a mute agent.

I will not claim those Tuesday rows as a client case study. They are the routing I am running on my stack. Your eval, your invoice, your hop.

---

What will I test this week versus what I will not claim? #

I will test pins, cliffs, and attended computer use. I will not claim bake-off winners, ROI, or access I do not have. Launch week is when blogs invent scores. I will not.

What I will test:

Test	Pass	I will not write
Astra coding hop after Sol retries	Patch lands, I can name the 272K math	"Astra is cheaper than Sol"
Astra computer use on a disposable VM	Four checks above, me in the chair	A client OS takeover
Fable promote after Opus high fails	Eval flips, hop is logged	"Fable replaces Opus"
3.8 Flash ID + thinking_level swap	Payload stops sending thinking_budget	"3.8 is the new Gemini flagship"
3.7 stay on a cheap loop	Tokens/task did not jump	Delete 3.7 from production
Grok 4.6 under 200k	Live sheet matches the $2 / $6 card	An invented Fast ID
Cursor Task slug	Job finishes, Task dies	That the Task is a teammate

What I will not claim, even if a chart is pretty:

• A studio-wide default of gpt-6-astra.
• A Cursor default of claude-fable-5-1.
• Mythos 5.1 or 3.8 Flash Cyber as something I can call.
• Daybreak extras as access I already have.
• Unattended computer use as a standing product.
• A dollar-per-draft winner across Astra, Fable, 3.8, and 4.6 without a shared rubric and a logged invoice.
• Any exploit path, "summary" of an exploit path, or reconstructed system-card attack.

If a number is not on an official page I linked, or on my invoice, it does not go on this board as a fact. Estimates stay labeled. Vendor benches stay vendor benches.

---

What do I pin in Cursor, n8n, and Antigravity today? #

A short pin list. No architecture rewrite. I have watched launch-week "upgrades" turn a working Sol node into a $50-output heater with Fast mode left on.

Day-of list:

1. Inventory every model string. Cursor rules, n8n HTTP nodes, MCP server configs, Antigravity managed agents, batch jobs, eval harnesses. Write the current ID next to the job, not next to the vendor.
2. Add Astra as a flagship lane, not a default. Hard OpenAI coding and attended computer use. reasoning.effort at high on those routes. Leave passing volume on 5.6.
3. Leave Opus 5 as the Claude default. Add claude-fable-5-1 as a named promote with a reason you can audit. Do not rename the default.
4. Swap Google coding/agent IDs to gemini-3.8-flash. Replace thinking_budget with thinking_level. Leave 3.7 on cheap loops. Accept 3.8 as the Antigravity default.
5. Pin grok-4.6 where the caller is Grok. Watch 200k. Do not invent a Fast ID. Do not write the Cursor Task slug into an API node.
6. Keep computer use attended. Responses + human on Astra. Preview + sandbox on 3.8 Flash. Nowhere for exploit-shaped work.
7. Refuse the SKUs you do not have. Mythos 5.1. 3.8 Flash Cyber. Daybreak extras you cannot call.

Prompt I paste into Cursor when I want the board applied without a creative restack:

Apply this routing board. Do not invent IDs.

- Hard OpenAI coding + attended computer use: gpt-6-astra, reasoning.effort=high. Do not set max on classifiers. Do not enable Fast studio-wide.
- OpenAI volume that already passes: keep GPT-5.6 Sol / Terra / Luna.
- Claude default complex: claude-opus-5. Promote to claude-fable-5-1 only after Opus at high effort fails the eval.
- Claude volume: claude-sonnet-5. Claude cheap: claude-haiku-4-5. Fable fallback: claude-opus-4-8.
- Google coding/agents/Antigravity: gemini-3.8-flash. thinking_level LOW on classify, MEDIUM or HIGH on agents. No thinking_budget. Keep gemini-3.7-flash on cheap loops.
- Grok callers: grok-4.6. Stay under 200k unless I approve the double. No grok-4.6-fast. Cursor Task slug cursor-grok-4.6-xhigh-fast is a disposable subagent, not a teammate.
- Do not route to claude-mythos-5-1 or Gemini 3.8 Flash Cyber.
- Computer use stays attended. No exploit write-ups.

That prompt is the board in a form a model can follow. It is not a new studio.

If a request is really "should we spend Astra tokens on this workflow," I run the ROI habit: name the hours, name the token path, name the failure cost. The ROI post is the longer version. This page is the September model inputs.

---

If your stack still treats Sol as the flagship, Fable as the new Opus, 3.8 as "the Gemini," or Grok as one noun, that is the work. Book an AI automation strategy call and I will seat the September board on your actual jobs — Astra for hard coding and attended computer use, Fable only after Opus 5 fails, 3.8 Flash for cheap volume, Grok 4.6 under 200k — or we scope a custom agent that routes by job instead of by the newest noun. I have done this across 600+ automations with 500+ still live. The win on a ship week is almost never more model. It is the right ID on the right loop before the invoice teaches you the difference.

---

Frequently asked questions #

Which frontier model should an operator use in September 2026? #

None of them as a single default. Pick by job: Astra for hard OpenAI coding and attended computer use, Opus 5 first on Claude with Fable 5.1 after a high-effort fail, 3.8 Flash for cheap volume and Antigravity, Grok 4.6 under 200k for Grok-native and Cursor subagents. The board above is the answer. One string is a keynote.

Which model should I use for coding this week? #

gpt-6-astra for hard OpenAI / Codex work, claude-opus-5 for Claude work that already passes, claude-fable-5-1 only after Opus at high effort fails, gemini-3.8-flash for Antigravity and Google coding agents, grok-4.6 when the caller is already Grok. DeepSWE is a photo finish on OpenAI's own September 3 table. Do not delete Sol, Opus, or Flash because Astra shipped.

Which model should I use for computer use? #

Astra on the OpenAI lane, attended, on Responses. Gemini 3.8 Flash on the Google lane as Preview, sandboxed. Computer use bills extra on Astra. Preview is not a standing OS write path on Gemini. I will not invent a Fable clicker, and I will not write exploit copy from a system card.

Which model should I use for cheap subagents? #

gemini-3.8-flash for cheap-but-strong Google loops, gemini-3.7-flash for high-QPS work that already passes, grok-4.6 under 200k or cursor-grok-4.6-xhigh-fast in Cursor, claude-haiku-4-5 for Claude glue, GPT-5.6 Sol / Terra / Luna for OpenAI volume. Do not put Astra or Fable on a classifier. Cheap is a seat.

Does GPT-6 Astra replace GPT-5.6 Sol? #

No. Sol, Terra, and Luna stay the cheaper OpenAI stack. I move hard computer-use and long Codex jobs and leave passing volume on 5.6 — OpenAI did not ship a GPT-6 cheap trio on September 3. Details live on the Astra spec card.

Should I switch from Claude Opus 5 to Fable 5.1? #

No. Promote when Opus at high effort still fails. That is Anthropic's own instruction in the Fable 5.1 overview and the rule I wrote in the Fable operator stack. A switch hides the jobs Opus already passed and doubles input versus Opus 5.

Is Gemini 3.8 Flash a flagship? #

No. It is the current Flash. Swap the ID, drop thinking_budget for thinking_level, keep 3.7 on cheap loops, and treat 3.8 Flash Cyber as Fairwind-only. I already wrote the operator swap. 3.1 Pro stays the preview reasoner.

Is Grok Bot a model I pin in Cursor? #

No. Grok Bot is a product, not a model ID. Pin grok-4.6 on the API and cursor-grok-4.6-xhigh-fast on a Cursor Task. I already wrote the Extra High Fast pin. This board stays on IDs.

What happens if a Grok 4.6 prompt crosses 200k tokens? #

The whole request doubles on the card I budget against — not just the tokens over the line. Docs give the 500k window and the $2 / $6 headline. Confirm the live sheet before you pin a 400k loop. Split the prompt or stay under 200k unless I approve the cliff.

What happens if a GPT-6 Astra prompt crosses 272K tokens? #

The full request prices at 2x input and cache and 1.5x output. A 400k dump is not "a little more." It is a different invoice. Fast is a separate 2x tax. Numbers are on the API model page.

Do I have Claude Mythos 5.1? #

No. Mythos 5.1 is invite-only under Project Glasswing. Same weights as Fable 5.1, different safeguards, not a public default. That is not the April 2026 Mythos post. I do not route to claude-mythos-5-1.

Is Gemini 3.8 Flash Cyber available on the Gemini API? #

No. Cyber is Fairwind-only. The public ID I will pin is gemini-3.8-flash. If a client forwards the Cyber paragraph, the answer is we do not have Fairwind. See the 3.8 Flash swap.

What reasoning effort should I set on Astra? #

medium or high on work I will read. xhigh or max only on a named job with me in the session. Never max on a classifier. OpenAI's charts were run at maximum effort. That is a scoreboard, not a standing-agent setting.

Does this replace the May 2026 frontier comparison? #

It updates the names. It keeps the method. The May pillar still tells you to compare context, price, and the job. The IDs in that piece are dated. Use this board for September 2026.

---

If your model picker is still one dropdown and a vibes argument, that is the signal. Book the AI automation strategy call. I will not pick one frontier model for your whole company. I will pin the board.