Claude API+25 Positive

Operator Insights

What are teams learning from building on Claude API and running it in production?

7 accounts: 1 failure, incident or rollback, 1 left it or turned it down, 2 run it beside another tool and 3 adopted it

Failures, incidents and rollbacks1

Pramod KalipatnapuCofounderRevefi
Engineering team. Raden data agent serving enterprise customers.
Revefi blog, published Sep 23, 2026Read the write-up (opens Revefi blog in a new tab)
Observation
At Revefi, one timestamp at the top of the prompt was enough to break caching for an agent.
Takeaway
Order every prompt from stable to volatile, and alert on cache hit rate per agent, because a broken cache fails silently and shows up only on the bill.
What happened
An upgrade added user context and a per-call timestamp near the top of Raden's system prompt. Revefi then saw its agent spend climb and its cache hit rate collapse, since nothing after the timestamp could be read from the cache. Kalipatnapu and Istasis Mishra moved volatile content to the end of the prompt. Cache reads rose to 99% of input tokens the week the fix shipped, and the post reports a 90% cut in the agent's spend.
“Earlier this year, we had to quickly reduce our data spend. We had inadvertently introduced a change that made our costs spiral out of control!”
Pramod Kalipatnapu and Istasis Mishra
- Public engineering write-up
- Spend and cache metrics reported
Deployment details
Engineering team. Raden data agent serving enterprise customers.
- Public engineering write-up
- Spend and cache metrics reported
Team
Revefi engineering
Workload
Long-running data agent, dozens of model turns per request
Product version
Claude API prompt caching, LangChain middleware
Measurement
Cache reads as a share of input tokens in the Anthropic Console; agent spend

Left it or turned it down1

Bruno ŠkvorcStaff Software EngineerLindy
Lindy, an AI assistant company running managed agents for email, calendar and meeting work.
Lindy engineering blog, published Jun 24, 2026Read the write-up (opens Lindy engineering blog in a new tab)
Observation
Lindy dropped Claude Sonnet for most managed-agent traffic and moved it to DeepSeek v4 Flash, cutting inference cost on those routes by about 90%.
Takeaway
Price every route against an open model on your own evals; for high-volume assistant traffic, Lindy found one that held quality at a tenth of the cost.
What happened
Lindy dropped Claude Sonnet and Gemini as the default for most managed-agent traffic in May and June 2026 and moved those routes to DeepSeek v4 Flash on Atlas Cloud. It ramped DeepSeek to 100% of target traffic only while offline evals, online evals and retention held, and inference cost on those routes fell by about 90% (Jun 24, 2026). Sonnet still runs when a user selects it. The post does not say whether Lindy reached Claude through the Claude API or a cloud.
“We moved most managed model traffic from Claude Sonnet and Gemini paths to DeepSeek v4 Flash on Atlas Cloud. Inference costs on those routes fell by about 90%.”
Bruno Škvorc
- Public engineering write-up
- Cost change reported
Deployment details
Lindy, an AI assistant company running managed agents for email, calendar and meeting work.
- Public engineering write-up
- Cost change reported
Team
Lindy engineering
Duration
Ramp over May and June 2026
Workload
Most managed-agent model traffic
Model
DeepSeek v4 Flash in place of Claude Sonnet and Gemini
Measurement
About 90% lower inference cost on the migrated routes (Jun 24, 2026)

Run it beside another tool2

Siddharth KodwaniSoftware Engineer, GenAI PlatformDoorDash
DoorDash's GenAI platform team, serving product teams across the company through one LLM Gateway.
InfoQ, published Oct 3, 2026Watch the talk (opens InfoQ in a new tab)
Observation
DoorDash runs the Claude API with Amazon Bedrock as its fallback, and OpenAI with Azure, behind one gateway that bills each department.
Takeaway
Buy Claude on two routes, the Claude API and a cloud, before you need the second one, and put cost attribution in the gateway from the first team onward.
What happened
DoorDash put OpenAI, Anthropic and Gemini models behind one API and SDK so product teams could try a new model with one change. For reliability each provider got a fallback: the Claude API endpoint falls back to Amazon Bedrock, and OpenAI to Azure (InfoQ, Oct 3, 2026). As usage grew, finance asked who was spending what on shared keys, so the gateway added workspaces with cost attribution per department. The team also names rising cost and model deprecations that force migrations across hundreds of use cases.
“We actually had to go understand what the fallback should be, what kind of routing should be. At one point, I remember we had OpenAI endpoint, but we also had a fallback on Azure endpoint. Or, we had Claude model API endpoint, then we had a fallback on Bedrock, so that the services can rely on us from the reliability point of view. We went in that direction. We quickly realized that as the usage is going up, then we need a way to track, because now finance got involved. Like, you guys have one API key for multiple teams, how do I track who is using what? We quickly realized we need workspaces where we have cost attribution. We can actually drill down on who is using how much. We can attribute cost back to those departments. That's what basically we did with LLM Gateway.”
Siddharth Kodwani
- Conference talk (InfoQ)
- No metrics reported
Deployment details
DoorDash's GenAI platform team, serving product teams across the company through one LLM Gateway.
- Conference talk (InfoQ)
- No metrics reported
Team
DoorDash GenAI Platform
Workload
Hundreds of product use cases behind one LLM Gateway
Product version
Claude API with an Amazon Bedrock fallback; OpenAI with an Azure fallback
Model
OpenAI, Anthropic and Gemini models

ReplitNov 21, 2024Lenny's PodcastReplit's agent uses Claude Sonnet for coding and OpenAI models in other roles of the same multi-agent system.
Amjad MasadCo-founder and CEOReplit
Replit Agent, a multi-agent system with manager, editor and critique models.
Lenny's Podcast, published Nov 21, 2024Listen to the episode (opens Lenny's Podcast in a new tab)
Observation
Replit's agent uses Claude Sonnet for coding and OpenAI models in other roles of the same multi-agent system.
Takeaway
Assign models by role inside one agent, and keep each provider swappable as the coding lead changes.
What happened
Masad called Anthropic's Sonnet the most important model Replit uses and the one it uses for coding, with OpenAI models alongside it for other roles in the multi-agent system, such as critique (Lenny's Podcast, Nov 21, 2024).
“The most important model that we use is the Sonnet model from Claude, from Anthropic, and it is the best model at coding. So that's the model we use for coding, but we use models from OpenAI as well because a multi-agent system.”
Amjad Masad, at 32:51
Podcast episode (Lenny's Podcast)
No metrics reported
Deployment details
Replit Agent, a multi-agent system with manager, editor and critique models.
Podcast episode (Lenny's Podcast)
No metrics reported
Team
Replit Agent
Workload
Coding agent with manager, editor and critique roles
Model
Claude Sonnet for coding; OpenAI models for other roles (Nov 2024)

Adopted it3

ProjectDiscoveryApr 10, 2026ProjectDiscovery blogProjectDiscovery runs its Neo security agents on Claude with prompt caching, which saved 59% of LLM cost and 70% over its last 10 days.
Parth MalhotraAuthor of ProjectDiscovery's engineering postProjectDiscovery
Neo, ProjectDiscovery's autonomous security testing platform, whose multi-agent workflows take 20 to 40 or more LLM steps per task.
ProjectDiscovery blog, published Apr 10, 2026Read the write-up (opens ProjectDiscovery blog in a new tab)
Observation
ProjectDiscovery runs its Neo security agents on Claude with prompt caching, which saved 59% of LLM cost and 70% over its last 10 days.
Takeaway
On long agent loops, cache placement is the largest cost lever; measure saving against the same tokens at full price from your own bill.
What happened
ProjectDiscovery runs Neo on Claude with Anthropic prompt caching at three cache breakpoints. Against the same token volume at full input price, caching saved 59% of LLM cost overall, 66% after the team reordered its prompts and 70% over the last 10 days, with 9.8 billion tokens served from cache (Apr 10, 2026).
“Overall, caching saved 59% on LLM costs compared to what the same token volume would have cost at full input rates. Post-optimization that number is 66%, and the last 10 days are at 70%.”
Parth Malhotra
Public engineering write-up
Cost saving measured from actual spend
Deployment details
Neo, ProjectDiscovery's autonomous security testing platform, whose multi-agent workflows take 20 to 40 or more LLM steps per task.
Public engineering write-up
Cost saving measured from actual spend
Team
Neo engineering at ProjectDiscovery
Workload
Multi-agent security testing, 20 to 40 or more LLM steps per task
Product version
Claude API prompt caching with three cache breakpoints
Measurement
59% saved overall, 66% after optimization, 70% over the last 10 days (Apr 10, 2026)
BarclaysOct 1, 2026AnthropicAt Barclays, Claude's first bank-wide workloads were knowledge search and email triage.
Craig BrightGroup Co-Chief Operating OfficerBarclays
Organization-wide. Barclays UK and Global Markets. Live since 2025.
Anthropic, published Oct 1, 2026Read the case study (opens Anthropic in a new tab)
Observation
At Barclays, Claude's first bank-wide workloads were knowledge search and email triage.
Takeaway
Retrieval and triage workloads with checkable answers can carry a regulated rollout at volume, provided governance and human oversight are in place before scale.
What happened
Barclays' Colleague Knowledge Assistant, built on Claude with retrieval-augmented generation, has been live since 2025. More than 16,000 colleagues use it, and it has handled over one million searches for staff serving over 20 million UK retail customers. In Global Markets, Claude models classify, enrich, and route about 120,000 client emails a day. Barclays applies governance, security controls, and human oversight to each use case.
Vendor case study
Usage figures reported by Anthropic and Barclays
Deployment details
Organization-wide. Barclays UK and Global Markets. Live since 2025.
Vendor case study
Usage figures reported by Anthropic and Barclays
Team
Barclays UK and Global Markets
Duration
Knowledge assistant live since 2025
Workload
Retrieval-augmented knowledge assistant and client email triage
Measurement
Adopting colleagues, searches handled, emails processed per day
PostHogNov 25, 2025PostHog blogAt PostHog, model upgrades changed the agent more than harness work did.
Michael MatlokaAuthor of PostHog's agent write-upPostHog
Product team. 1 year. PostHog AI, the agent built into PostHog.
PostHog blog, published Nov 25, 2025Read the write-up (opens PostHog blog in a new tab)
Observation
At PostHog, model upgrades changed the agent more than harness work did.
Takeaway
Expect each model generation to change what your agent can do, and keep your own calling code thin so you can adopt the next model quickly.
What happened
PostHog spent a year building PostHog AI from a one-tool hackathon prototype into an agent that runs dozens of tools in a single loop. Matloka names two model step changes: OpenAI's o4-mini for reasoning over complex queries, and reliable tool use with the Claude 4 family. The latter let the agent use more tools without going off track. The core loop runs on Claude Sonnet 4.5, and the team stays off LLM frameworks.
“We're currently using Claude Sonnet 4.5 for the core agent loop, as it hits a sweet spot of quality, speed, and cost…”
Michael Matloka
Public engineering write-up
Model stated
No metrics reported
Deployment details
Product team. 1 year. PostHog AI, the agent built into PostHog.
Public engineering write-up
Model stated
No metrics reported
Team
PostHog AI team
Duration
1 year
Workload
In-product analytics agent with dozens of tools
Product version
Claude API tool use, single agent loop, no LLM framework
Model
Claude Sonnet 4.5 for the core loop

Failures, incidents and rollbacks1

At Revefi, one timestamp at the top of the prompt was enough to break caching for an agent.

Left it or turned it down1

Lindy dropped Claude Sonnet for most managed-agent traffic and moved it to DeepSeek v4 Flash, cutting inference cost on those routes by about 90%.

Run it beside another tool2

DoorDash runs the Claude API with Amazon Bedrock as its fallback, and OpenAI with Azure, behind one gateway that bills each department.

Replit's agent uses Claude Sonnet for coding and OpenAI models in other roles of the same multi-agent system.

Adopted it3

ProjectDiscovery runs its Neo security agents on Claude with prompt caching, which saved 59% of LLM cost and 70% over its last 10 days.

At Barclays, Claude's first bank-wide workloads were knowledge search and email triage.

At PostHog, model upgrades changed the agent more than harness work did.