Prepared for — a worked example of the account dossiers I build.
Cognition AI (Devin) & the Claude Expansion Blueprint
How I would run this account as Anthropic's enterprise AE. Devin already runs on Claude — Cognition said so when it rebuilt Devin for Sonnet 4.5. So the play is not migration; it is consumption depth and defence: deeper prompt caching, a procurement route through the AWS relationship Cognition just signed, and a credible answer to the in-house SWE-1.5 / SWE-2 models that compete for the same tokens.
Thesis: at Devin's scale, the cache-read line is the margin. Every point of cacheable-prefix share is worth more than any list-price discount, and it is the one lever Cognition's own models cannot copy without rebuilding Claude.
Current vs. optimised Claude spend — and what it does to gross margin per ACU
Model: current Sonnet tier at $2 / $10 per MTok with 5% cache reads, verified from Anthropic's pricing page. Three presets, two headline sliders, everything else under Advanced. The URL updates as you move sliders, so any scenario is shareable.
Advanced assumptions
Batch (50% off) is only applied to the batch-eligible share — evals, re-indexing, nightly review — because batch jobs are asynchronous. Interactive agent loops stay on the standard tier. Thinking tokens are billed as output.
Where the dollars go uncached input · cache writes · cache reads · output · thinking
What moves the number one-at-a-time sensitivity around the optimised scenario
Green = cheaper than the optimised baseline, coral = more expensive. If cache-hit share is on top, the conversation is about prompt architecture, not list price.
Build vs. buy — the real competitor is internal
Cognition trained SWE-grep and SWE-1.5 on GB200s and serves them on Cerebras at 950 tok/s. Anthropic's defence is not "Claude is smarter"; it is a per-loop cost that a hybrid stack can match while keeping Claude on the long-horizon reasoning.
Claude only optimised Sonnet
- per loop at the optimised scenario above
- single vendor, one prompt stack, 4 cache breakpoints
Hybrid Haiku 5.5 retrieval + Sonnet
- Haiku-class sub-agent ($0.10 / $0.50) does the SWE-grep job; main context shrinks 40%
- saves —/mo vs. Claude-only
In-house SWE-1.5 / SWE-2 on Cerebras
- placeholder — no public serving-cost figure exists; the gap to discover is —/mo
- Cognition itself positions these as complementary, not replacements
Volumes are scenario anchors, not Cognition disclosures. Revenue per loop derives from the public $2.25/ACU launch price and an editable loops-per-ACU assumption. The in-house serving cost is a stated placeholder. The point of the model is the shape of the economics and the questions it generates — see MEDDPICC.
Three hypotheses I would bring to the first technical call
Written as hypotheses to validate with Cognition's team, not as a lecture. Each ends with what I would want to check.
One deliberate pass beats N sampled ones
Agent harnesses historically bought reliability with self-consistency: sample a plan 3–5 times, vote, re-read the repo each time. Extended thinking moves that deliberation inside a single turn, with a budget the harness controls. The context is read once; the thinking tokens are billed as output.
Napkin math at 60K context: 3 samples re-read 180K input tokens; one pass with an 8K thinking budget reads 60K and writes 8K. At $2 / $10 that is roughly $0.36 vs $0.20 per planning step before caching — and with 90% cache reads the single-pass figure drops further because the one read is mostly cached.
- Which task classes still use multi-sample planning, and what is their share of loops?
- Is the thinking budget fixed or adaptive — and does it correlate with Junior Developer Eval difficulty?
- Does the 200K cap stay necessary once thinking replaces re-reads?
Sources: Cognition on Sonnet 4.5 (+18% planning, +12% JDE) · thinking tokens billed as output
| Model (SWE-bench Verified) | Resolved | Date |
|---|
Leaderboard excerpt from swebench.com, fetched 2026-10-13. Cognition stopped reporting SWE-bench in 2024 and runs its own Junior Developer Evals — so this table frames the market, not Devin.
Devin's prompt stack maps onto four cache breakpoints
Caching rewards stable prefixes. A coding agent's prompt is unusually prefix-stable: system instructions and tool schemas never change inside a session; the repo map and DeepWiki-style summaries change rarely; only the latest turn is new. Anthropic allows up to 4 explicit breakpoints with a 20-block lookback, 512-token minimum on the current models, and a TTL that refreshes on every read.
- Where does session state currently live — in the prompt (cacheable) or re-fetched per turn (not)?
- Typical gap between turns while Devin waits on CI or a human: under or over five minutes?
- Does parallel-Devin fan-out share a prefix, so one write serves many readers?
Sources: prompt caching docs · Devin 2.0: Devin Wiki, Devin Search, parallel Devins
Buy Claude where Cognition already sells Devin
Cognition's production environment is AWS (its own security docs say so), it signed a multi-year Strategic Collaboration Agreement with AWS in September 2026, and Devin is sold on AWS Marketplace. Bedrock has price parity with Anthropic direct — Sonnet 5.5 at $2 / $10, cache reads at $0.10 after the October 7 cut — so routing a Claude commit through the AWS relationship costs nothing on list price and can remove a vendor security cycle.
- Direct Anthropic API, Bedrock, or both today — and which one carries the volume?
- Is there an AWS commitment the SCA sits under, and does Marketplace spend count toward it?
- Batch on Bedrock is N/A for the 5.5-generation models: does any batch-eligible workload need to stay direct?
Sources: AWS–Cognition SCA (2026-09-15) · Devin security docs · Bedrock pricing
What I know, what I think, and what only discovery can answer
Highlighted cards are the front-half pillars an AE must own before a first meeting: Metrics, Economic Buyer, Decision Criteria, Identify Pain, Champion. Every claim links to its source.
The method is the product. Cognition is the worked example.
This page took one working session with an AI-native research pipeline. The same pipeline produces a dossier like this for any account on a target list — swap one content file, re-verify, ship.
Ground truth first
Live-fetch vendor pricing pages, the account's blog, careers board, security docs and press. Every number carries a URL and a fetch date.
Network evidence
Public DNS over HTTPS plus ARIN RDAP to confirm the production cloud and SaaS stack — no scanning, no login, nothing non-public.
Evidence ledger → badges
Each claim is graded Validated / Hypothesised / Unknown. The MEDDPICC badges are generated from the ledger, not from opinion.
Economics with unit tests
The calculator is a pure-function module with golden tests, including the hand-computed scenario and overflow checks at 10M loops × 150K tokens.