
Silicon 2 Turned Up in a Devin Model Table at Fable 5.1's Exact Price — What the Entry Says, and What It Doesn't
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0717 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 87 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 133 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 72 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 127 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 65 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 44 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 193 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 148 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 123 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 219 tok/s
A model entry called Silicon 2 has been showing up in builds of Devin Desktop, priced at $10 per million input tokens and $50 per million output tokens. Those are the exact numbers on Claude Fable 5.1's rate card, which has not moved since September 1, and they are 25% above what Claude Fable 5.2 charges for the same million tokens today. Nobody at Anthropic has said the word. Cognition, which ships Devin, has not said it either. The entry sits in a third-party tool's model table with the maker listed as unknown, and the people who noticed it disagree only about which Anthropic model it might become. That is the entire public record as of November 17, 2026: one screenshot of a model picker, a handful of posts, and a price that looks copied rather than set. Everything below is written to that standard — what was seen, what it implies, what would settle it, and what you should do in the meantime, which is very little.
Two entries, one of which got the attention
The screenshot that started this is a tracker card, not a Devin system prompt or a leaked config file. It shows two stealth entries logged at the same minute — 1:49 PM on November 16, 2026 — both attributed to "Devin Desktop build" and both labelled "Anonymous name — maker not confirmed." The tracker identifies itself as Model Scout and links a Discord community rather than a vendor source, so treat the provenance as "someone with a Devin build screenshotted a model table," which is a real thing to have and a weak thing to have at the same time.

• silicon-2 — $10 input and $50 output per million tokens, four reasoning tiers listed as high, low, medium and xhigh.
• roadrunner — $1.50 input and $9 output per million tokens, four tiers listed as high, low, medium and minimal.
The second entry is the one that makes the first legible. A stealth model table with two rows is a vendor testing two things at once, and the two things are at wildly different price points: one at a tenth of the other's input rate, one sitting exactly on Anthropic's old frontier price. The cheap entry has attracted almost no discussion, presumably because nobody has a story to hang on it. The expensive one has a story because the expensive one is priced like a Claude.
The chain of custody is short and worth spelling out. Chetaslua posted the screenshot on X at 11:02 UTC on November 16 with the caption "New stealth model spotted inside Devin Build — Silicon 2 — it is a new fable class model as you can guess with pricing," adding "Anthropic" and "Sits in the Claude block. Uses Fable 5.1's $10/$50 as a placeholder price." Chubby, posting as @kimmonismus, amplified it two and a half hours later with the framing this article inherited: reportedly surfaced in Devin, suspected Anthropic, sits among the Claude models, placeholder pricing matching Claude Fable 5.1. By the time we pulled the thread it had 375 likes and 75 replies, which is the size of a signal that has circulated inside one community and has not yet been picked up by anyone with a desk at Anthropic.
One prior item belongs in the file and cannot be verified. A trade write-up dated November 8, 2026 carries the headline "Devin CLI's New Build Lists a 'Silicon' Model in Five Tiers." The page returned a bot challenge on every fetch attempt we made, so we have its headline and its date and nothing else — not the tier names, not the prices, not whether it is describing the same entry. If it is, the model now styled Silicon 2 has been visible in Devin's interface for at least eight days, and "2" would be a version number on a codename already in use rather than a new sighting. We can't confirm that, so we're not building on it.
The price is the whole argument, and it points backwards
Strip the speculation out and one fact carries the piece: Claude Fable 5.1 costs $10 per million input and $50 per million output. That is Anthropic's published rate on its own models overview, still live today, next to Claude Fable 5.2 at $8 / $40, Claude Opus 5.5 at $4 / $20 and Claude Sonnet 5.5 at $2 / $10. A stealth entry landing on $10 / $50 is not a coincidence you can wave away — it is the same card, to the dollar, as one specific model in the current lineup.

But read which model. When Anthropic shipped Claude Fable 5.2 on November 9, it cut the frontier price by 20%, from $10 / $50 to $8 / $40, and dropped cache reads to 2.5% of base input on both the 5.2 and 5.1 generations. If a new Anthropic flagship were being staged inside Devin this week, the number you would expect to see in that table is $8 / $40 or something below it — not a return to the price Anthropic just walked away from. A placeholder copied from the previous generation's rate card is exactly what an unannounced model looks like before its pricing is set: correct shape, stale value.
That reading cuts both ways for the Anthropic hypothesis, and it would be dishonest to present only the flattering half.
• For Anthropic — the four tier labels include xhigh. That word is not generic. Anthropic's effort parameter documents five levels, low, medium, high, xhigh and max, and xhigh is restricted to the Fable, Mythos and top Opus/Sonnet/Haiku 5.x families. A vendor with no exposure to Anthropic's API would not reach for "xhigh" as a tier name.
• Against a clean read — the same tier list is missing max, which Anthropic also offers on Claude Fable 5.2, and the sibling entry uses "minimal," which is nobody's effort level. Devin is evidently normalising model-specific settings into its own four-slot enum, so the tier labels are Devin's vocabulary, not Anthropic's, and should not be read as evidence of a specific underlying model.
The honest summary is that the entry is Anthropic-shaped and Anthropic-priced and could still be something else. Devin's own documentation says the tool supports Anthropic, OpenAI, Google, Cognition and several open-weight families — DeepSeek, Kimi, GLM — and that short names like opus, sonnet, swe, codex and gemini resolve to the newest version of each family. A major lab testing a launch partner's client before an announcement is the ordinary explanation. A competitor benchmarking against a frontier price is a possible one. A placeholder row that someone typed into a build and never deleted is not impossible, though $10 / $50 is a strange number to invent.
What "Silicon" would be a codename for, if it is one
Anthropic does not publish codenames and does not confirm them. What it does do, reliably, is let third-party tooling see a model shortly before the model exists publicly — in model pickers, in client builds, in configuration files that ship to users. That pattern is why a screenshot of a dropdown can be worth more than a hundred replies speculating about one. It is also why the same screenshot can be worth nothing: the gap between "a row exists in a client" and "a model is callable" has swallowed plenty of predictions.
The numbering is the part people keep getting wrong. "Silicon 2" does not imply a Silicon 1 that anyone has seen, and even if the November 8 headline refers to a "Silicon" entry, a 2 on an internal name tells you about the lab's internal iteration count, not about the public lineup. Anthropic's public naming has been a family word plus a version — Fable, Mythos, Opus, Sonnet, Haiku — and a new family word would be a bigger event than a new version. If Silicon is a family, the interesting question is not whether it beats Claude Fable 5.2 but what slot it occupies. The $10 / $50 price hints at the frontier slot; the absence of max in the tier list hints at a model that is not quite the same shape as Claude Fable 5.2. Neither hint survives contact with a real model card.
Five minutes of checks that would settle it
This is the part that most coverage of a sighting skips, and it is the part that makes a leak item useful rather than merely exciting. Every check below is something you can run yourself, and each one moves the question from "someone saw a row" to "the row is real."
• Anthropic's models overview — the page that lists every model with its API ID, context window, max output, per-million prices and retirement commitment. A shipping Anthropic model appears here, with a platform ID for Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. As of today the newest rows are Claude Fable 5.2 and Claude Mythos 5.2, both dated November 9, 2026.
• The Models API — every available model, its token limits, its capabilities object and its line field, queryable programmatically. An unannounced model does not appear; it also does not appear in any of the 6,023 entries in LiteLLM's public price and context map, or in the 324 providers tracked by models.dev, both of which we checked today and neither of which carries a Silicon entry.
• The price sheet — Anthropic's pricing page carries the cache-write and cache-read rates, the batch discount and the long-context rates, and it is where a new model's real economics show up first. Nothing named Silicon is on it.
• An independent leaderboard — Artificial Analysis lists a model when it can evaluate it, which requires a callable endpoint. A frontier-class launch that has been evaluated and not listed is a contradiction.
• The lab's own account — Anthropic's news index, dated. On November 9 it carried Claude Fable 5.2 and Claude Mythos 5.2. It carries nothing since.
All five are currently negative, which is the correct result for a model three weeks before an announcement and also the correct result for a model that does not exist. The point of the list is that it will stop being ambiguous the moment one of them flips, and the first flip is nearly always the models overview or a partner platform ID, not a post.
What it would actually cost, if it ships
Assuming for a moment that Silicon 2 is a real Anthropic frontier model and that the entry's price is not the placeholder everyone assumes, the economics are easy to state and hard to like. At $10 / $50 you would be paying 25% more per input token and 25% more per output token than Claude Fable 5.2 charges today for a model that may be its successor. Output dominates most agent bills, and the gap widens with every tool call: Claude Fable 5.2's documented tool-call behaviour in agentic work is already the expensive part of a session, and a 25% higher output rate compounds across every turn the agent takes to converge.
Two things would change that arithmetic, and neither is knowable yet. The first is cache pricing. Anthropic put cache reads at 2.5% of base input on the Fable 5.2 and Fable 5.1 generations; a new model can hold that, improve on it, or lose it, and for anything that re-sends a long system prompt on every turn the cache line matters more than the headline rate. The second is effort behaviour. If a model reaches the same result at medium that its predecessor needed at high, the per-token premium is a discount in practice. That is measurable per workload and completely unmeasurable from a screenshot of a dropdown.
The practical point for anyone planning a migration is that none of this is a reason to wait. Claude Fable 5.2 has been callable since November 9 at $8 / $40, with a 1M-token context window, 128K max output and a June 2026 knowledge cutoff. Claude Fable 5.1 is still on the sheet at $10 / $50 if a specific eval of yours prefers it, and Claude Opus 5.5 is at $4 / $20 for the workloads where the frontier premium does not pay back. A model whose maker has not confirmed it has no SLA, no rate card, no retirement date and no route. Betting a production path on it is a bet on a screenshot.
Where a sighting like this genuinely earns its keep is in the testing budget, not the production path. OrcaRouter carries Claude Fable 5.2, Claude Fable 5.1 and Claude Opus 5.5 on one endpoint at Anthropic's list price with zero markup, so a price change on the vendor's sheet is live on ours the same day rather than after a renegotiation. That makes the cheap version of the Silicon 2 decision available now: run your own eval against Claude Fable 5.2 and Claude Fable 5.1 at their real rates, keep the prompt and tool schema you would use for the new model, and have a fallback configured. If an announcement comes and the model behaves differently from what that eval predicted, failover moves the traffic without a code change. If it never comes, you lost an afternoon and you now know what the current frontier costs you per task — which is the number you actually needed.

Questions worth asking that a spec sheet will not answer
Is there any reason to hold a model migration for Silicon 2?
No, and the reason is structural rather than a judgment about this particular sighting. A migration decision needs a rate card, a context window, an effort model and a failure mode, and an unconfirmed entry has none of them. What a sighting like this one does justify is keeping the migration cheap: one endpoint, two candidate models, a routing rule you can repoint. If Silicon 2 turns out to be real and turns out to be better on your workload, the swap should cost you an afternoon, not a quarter.
Does a stealth entry in a partner tool usually precede a launch?
Often enough that it is worth watching and not often enough to plan around. Third-party clients do see models early — that is how most recent frontier launches were first spotted — but the same channel also carries entries for models that ship months later, under a different name, at a different price, or not at all. The entry is a leading indicator that a lab is doing partner work. It is not a date.
What to watch, and what would change this piece
The first thing to watch is not a post. It is a row: a new ID on Anthropic's models overview, or a partner platform ID alongside the existing ones, or the same name surfacing in a public price map. Any of those within the next few weeks would turn this from a sighting into a launch, and the price on the row would tell you immediately whether $10 / $50 was a placeholder or a decision. The second is Devin's own model documentation, which currently lists Anthropic, OpenAI, Google, Cognition and the open-weight families and does not mention Silicon at all; a change there is the partner confirming something the lab has not.
Until one of those happens, the accurate sentence is the boring one. A model entry called Silicon 2 exists in a third-party build, it is priced like Claude Fable 5.1, its maker is listed as unknown, and the people who found it are guessing. That is worth knowing, because the guess is well-founded and the price is oddly precise. It is not worth acting on, because Claude Fable 5.2 is real, costs less per token than the entry claims, and is available today at the rate Anthropic publishes — which is the only price anyone can actually be charged.
Compared in this article1
Detected from this article · Benchmarks: Artificial Analysis · updated daily
