Leaving direct Anthropic access
what Back Market measured, and what the other sources say

Nicolas M opens on what he calls « là où ça fait mal », where it hurts. In May 2026, about 230 to 240 developers consume 57 k$ of Anthropic API, through Claude Code alone. In June, the same perimeter reaches 87 k$.

He speaks as the person in charge of the rollout: « Imaginez ceci, 250 développeurs pendant 1 an qui peuvent utiliser Claude Code sans aucune limite de budget avec une facturation à l'usage. » Picture 250 developers for a year, using Claude Code with no budget limit and usage-based billing. His assigned role is to deploy the tool and train people, while the spend climbs.

The decision is taken to leave direct Anthropic access. OpenRouter usage starts on 18/06/2026, and the rollout becomes general at the end of July.

Nicolas M shows the 2 consoles side by side. The Anthropic column goes dark while the OpenRouter column rises, over the same user population.

Enlarge ↗

Asked whether AI use had stopped, he answers: « Est-ce que vous avez arrêté d'utiliser l'IA ? Non, tous ces gens-là ont été transférés sur OpenRouter. » Did you stop using AI? No, all those people were moved to OpenRouter. The Anthropic column falls because the usage moved to OpenRouter.

September's review is not the first episode. On 22/07/2026, Nicolas M had already explained the mechanism, and above all what it costs in legal work. That is the part spend tables never show.

His commercial argument is plain: « vous en tant qu'entreprise vous allez avoir qu'un seul contrat avec Open Router et vous n'aurez pas forcément besoin d'avoir ensuite un contrat avec tous les fournisseurs qu'il y a derrière ». As a company you will have a single contract with OpenRouter, and you will not necessarily need a contract with every provider behind it.

Enlarge ↗

He draws a rule from it: « si vous routez vers 10 providers, vous avez 10 politiques de données à vérifier ». Route to 10 providers and you have 10 data policies to check. He says out loud that this concerns providers and not models, correcting a slip on his own slide.

The same episode defines, in passing, the vocabulary used from here on: a frontier model is one « parce qu'en fait c'est un modèle très puissant et qui ne peut pas tourner aujourd'hui dans votre infrastructure ou sur un provider, et c'est surtout un modèle qui est privé », a very powerful model that cannot run today in your own infrastructure or at a provider, and above all a private model.

Anthropic was then the only host of the Claude models. The move to OpenRouter does not change that. It does not open Claude to other hosts, it opens usage to other models.

In August, Nicolas M crosses 2 views of the same console, the split by amount spent and the split by token volume. A model that weighs more in tokens than in spend costs less per token than the average of the mix.

Enlarge ↗

GLM 5.2 is both the most used model and, in his words, « le modèle aussi qui va coûter le plus cher », the one that will cost the most, among those carrying the volume. Yet when the sort runs on money rather than tokens, Sonnet and Opus climb back despite their low volume.

Hence his conclusion: « les modèles d'Anthropic sont vraiment vraiment chers, même quand vous passez par OpenRouter », Anthropic's models are really expensive, even when you go through OpenRouter.

He takes care not to overread the disappearance of Opus: « le coût a été réduit pas parce qu'on s'en sert moins mais parce que simplement on a accès à plus de providers », the cost went down not because we use it less but simply because we have access to more providers. He also answers the productivity objection: « les utilisateurs n'ont pas perdu soudainement leur productivité ou leur usage », users did not suddenly lose their productivity or their usage.

No access to Back Market's console is needed. The OpenRouter catalog is public and requires no authentication, and his claim about the price of Anthropic models can be checked against it.

Back Market works with about ten providers, including Microsoft Azure, Amazon Bedrock, Google, Fireworks and Mistral. Nicolas M notes that Mistral also hosts GLM 5.2: « vous pouvez aller taper sur l'infrastructure Mistral », you can go hit Mistral's infrastructure, to consume a Chinese model from European infrastructure.

His definition, for anyone who confuses the 2 levels: « un provider c'est comme un hébergeur internet qui aurait plein de sites internet, et chaque site internet en fait ce sont les fameux modèles », a provider is like a web host that has lots of websites, and each website is actually one of the famous models.

He describes the arbitration: « cette ventilation, elle est faite automatiquement par OpenRouter selon aussi le prix du token. Parfois c'est plus intéressant d'aller chercher chez Microsoft un accès à OpenAI que de l'avoir en direct, et parfois c'est l'inverse. », this allocation is done automatically by OpenRouter, also based on token price. Sometimes it is more advantageous to get access to OpenAI through Microsoft than to have it directly, and sometimes it is the other way around. The benefit he values is operational more than financial: « moi, en tant qu'utilisateur, j'ai pas besoin de m'ennuyer », as a user, I don't need to bother with it.

On OpenRouter's side, these constraints go through documented routing parameters. zdr: true restricts a request to endpoints with no data retention, data_collection: "deny" excludes providers that retain data. The only list restricts routing to named providers, within those allowed at the account level. This is the technical equivalent of Back Market's short list described in part I. By default, OpenRouter distributes load favoring the cheapest providers, weighted by the inverse square of their price. Source: provider routing documentation.

Making a workstation independent of a model provider takes more than a gateway URL. The article "Portability becomes a Scale concern" describes what survives a controlled migration from one provider to another.

His preferred metric sits in the top right of the OpenRouter console, the blended dollar per million. Between July and August, it falls 13 %.

Enlarge ↗

He reads it this way: « mes utilisateurs pour un million de tokens ont payé moins cher que le mois d'avant, et je suppose que du coup ils ont utilisé des modèles moins chers. Pour mesurer l'efficience d'une entreprise et de 245-250 utilisateurs, c'est plutôt un bon indicateur. », for a million tokens, my users paid less than the previous month, and I assume they therefore used cheaper models. To measure the efficiency of a company and of 245 to 250 users, it is a fairly good indicator.

The same blind spot exists at the task level. Inferya compares 2 models that each cost 0.07 $ per attempt on SWE-bench Verified, with the mini-SWE-agent harness: one solves 75.8 % of tasks, the other 9 %. Per task solved, the first works out to 0.092 $ per task and the second to 0.778 $, that is 8.4 times more.

A similar blended cost can therefore hide very different costs per task solved.

The FinOps Foundation formalizes these indicators, from cost per token to cost per inference and per API call, in its "FinOps for AI" guide, updated on 17/02/2026. The article "Mapping the token-reduction toolbox" applies the same caution to token-reduction tools. A shorter output is not enough to prove that a completed task costs less.

Nicolas M presents this drop as an inference, not as a measurement. The verb he uses is « je suppose », I assume.

The budget resets to zero at the end of the month. Asked whether 200 $ is enough, he answers: « Oui, les gens le prouvent aujourd'hui. », yes, people prove it today.

1. Self-served 20 $ extension

   Requestable from Slack at any time, « parce que c'est pas un humain qui va approuver votre demande », because it's not a human who will approve your request. It requires filling in a form with figures taken from the OpenRouter console: number of sessions, most expensive session, breakdown by model. It is granted only once: « si j'éclate ces 20 dollars, j'aurai pas le droit à 240 dollars, ça s'arrêtera là », if I blow through those 20 dollars, I won't get another 240 dollars, it stops there.

2. Project exception arbitrated by committee

   The manager fills in a Confluence form with an estimate of the amount needed. A committee, combining budget holders and product profiles, checks that the request « va générer du business », will generate business. Stated turnaround: about half a day. Result: a dedicated OpenRouter workspace, outside the individual allowance.

3. Dedicated keys for automated processing

   Non-human agents, pull request review, CI tasks, Slack bots, product features, have their own keys with an imposed model, and do not consume a developer's allowance.

The intent of the form is educational, not punitive: « c'est pour que les gens puissent comprendre un peu où est parti l'argent et qu'ils puissent se poser les bonnes questions, et que nous en fait on n'ait pas besoin de leur expliquer à chaque fois qu'est-ce que c'est qu'un modèle, qu'est-ce que c'est qu'une session. », it's so people can understand a bit where the money went and ask themselves the right questions, and so that we don't need to explain every time what a model is, what a session is. He adds that it is « super bien accepté par nos collaborateurs », very well accepted by our colleagues.

The guide AI Unit Economics distinguishes budget policies by who is spending. An agent, a CI task or an unsupervised service call for a strict cap, since there is nobody to read an alert. A developer in a session can receive an alert, then go through an approval.

He cites amounts tied to projects: « on sait qu'on a mis 1100 € sur ce petit projet là qui est important, et que ce hackathon là finalement a coûté que 38 dollars et pas 500 dollars », we know we put 1,100 € into that small but important project, and that hackathon actually cost only 38 dollars, not 500 dollars.

The same reasoning applies to automated agents, with an additional constraint: « si jamais un système commence à faire n'importe quoi et à dépenser plein d'argent, grâce à OpenRouter on peut très facilement contrôler ça, mettre de l'alerting dessus », if a system ever starts acting erratically and spending a lot of money, thanks to OpenRouter we can very easily control that, put alerting on it.

His conclusion deliberately widens the subject: « je sais que je parle beaucoup d'OpenRouter, mais au-delà d'OpenRouter, ce que je voudrais vous montrer c'est comment on fait pour que la valeur que génère l'intelligence artificielle, on s'assure que cette valeur est là » I know I talk a lot about OpenRouter, but beyond OpenRouter, what I want to show you is how we make sure the value generated by artificial intelligence is actually there. He adds a non-financial argument: « quand on voit des tokens et la finance, faut réfléchir que ça a un certain coût en énergie », when we look at tokens and finance, we need to remember that it has a certain energy cost.

Between 2025 and September 2026, 75 public videos mention OpenRouter, across French- and English-speaking technical channels. Most cite it in passing; 12 bring a fact, a figure or a firsthand account.

For an overview of the market, OpenRouter and a16z published on 04/12/2025 a « State of AI » report based on more than 100,000 billion tokens routed in one year. It finds that the adoption of open weight models has accelerated and that they now serve as building blocks for production applications.

The Claude Code Ultimate Guide already covered the question of gateways, but scattered across pages, and with a contradiction between 2 of them. One advised against pointing Claude Code at another backend in production, the other recommended exactly this mechanism.

Each lesson from this document is incorporated page by page, without repeating the same figure in 2 places.

2 other pages of the guide extend the topic without revisiting the Back Market case: AI Unit Economics, on cost per accepted task, and Team Metrics, on the metrics of teams working with agents.

Model substitution is the first thing to test. It needs neither a team gateway nor a spending cap. The test runs on a small, fixed batch of real tasks, 5 to 10 requests representative of everyday work, sent identically to 2 models.

The most direct test connects Claude Code itself to OpenRouter. The official integration guide fits in 3 environment variables, with no local proxy, to place in the shell profile or in the project's .claude/settings.local.json file.

export OPENROUTER_API_KEY="sk-or-..."
export ANTHROPIC_BASE_URL="https://openrouter.ai/api"
export ANTHROPIC_AUTH_TOKEN="$OPENROUTER_API_KEY"
export ANTHROPIC_API_KEY=""   # doit être vide, et non absente

If Claude Code was already connected to an Anthropic account, run /logout once and restart. The /status command should then display https://openrouter.ai/api on the base URL line. This URL takes no /v1 suffix, reserved for the OpenAI-compatible API, which triggers model-not-found errors.

For a controlled comparison outside Claude Code, the first step is to list the available identifiers rather than copy a name seen in a video. The catalogue is public and requires no authentication.

curl -s https://openrouter.ai/api/v1/models -o or-models.json

python3 - or-models.json <<'PY'
import json, sys
rows = json.load(open(sys.argv[1]))["data"]
for m in sorted(rows, key=lambda x: float(x["pricing"]["completion"])):
    p = m["pricing"]
    print("%-44s in=%7.2f $/M  out=%8.2f $/M" % (
        m["id"], float(p["prompt"]) * 1e6, float(p["completion"]) * 1e6))
PY

The response contains data, total_count and links; each entry carries id, pricing.prompt, pricing.completion and context_length. As of 16/09/2026, the catalogue holds 443 entries. Identifiers prefixed with a ~ are aliases that follow a family's latest version: convenient for exploring, best avoided for comparisons, since their target can change between 2 passes.

Second step, send the same request to both models. The following command reads the request from prompt.txt and saves each response to a separate file, from any working directory.

export OPENROUTER_API_KEY='sk-or-...'

for model in anthropic/claude-sonnet-5 z-ai/glm-5.2; do
  jq -n --arg m "$model" --rawfile p prompt.txt \
    '{model: $m, messages: [{role: "user", content: $p}]}' \
  | curl -s https://openrouter.ai/api/v1/chat/completions \
      -H "Authorization: Bearer $OPENROUTER_API_KEY" \
      -H "Content-Type: application/json" \
      -d @- > "reponse-${model//\//_}.json"
done

Third step, compare on verifiable criteria rather than impression: is the response usable without rework, does the code compile, do the tests pass. Then relate each pass's token consumption to the price noted in the first step.

Without this count, there is no way to know whether the cheaper model stays cheaper once rework is counted. The AI Unit Economics guide explains how to build this pairwise comparison, and what a small sample allows one to conclude.

Back Market's individual allocation rests on a documented capability of the management API. A key can carry its own credit limit.

On its Enterprise page, OpenRouter documents per-key credit caps with automatic reset, daily, weekly or monthly. One key per developer with a monthly reset therefore reproduces Back Market's 200 $ per month allocation, without in-house accounting.

For an individual test, reproducing the 3-question form depends on no API. After a week of use, answer for yourself: how many sessions, which session cost the most, and what is the split by model. Nicolas M credits it with a pedagogical effect, and it can be tested without a budget.

Enlarge ↗

The population figures vary within the video itself: the title announces 245 users, the introduction speaks of 250, the first Anthropic screen of 230 to 240. Nicolas M uses them interchangeably to refer to the same population. This document keeps his wording rather than harmonizing a figure he never settled on.

Blended cost

Average price paid per million tokens over a period, across all models. OpenRouter displays it on its console. At constant volume, it falls as users shift toward cheaper models.

Open weight

Model whose weights are published, so it can be hosted by any provider. Distinct from a proprietary model, accessible only at its publisher.

Provider

Host that serves a model. The same open-weight model can be served by several providers, at different prices.

Zero data retention

Commitment by a provider to retain no request data. Selection criterion cited to rule out otherwise well-ranked models. Abbreviated ZDR.

Conclusion

Moving to a gateway and capping spend per person are 2 separate decisions, which Nicolas M presents as joint. The gateway opens access to other models and breaks the spend down by model; the cap limits what each developer commits. The video does not allow saying which share of the drop belongs to each.

The cheapest transferable mechanism is the 3-question form required before the 20 $ extension. It forces the developer to read their own console before spending more. It needs neither a gateway nor a budget, and can be tested alone, on one's own usage.

Dated sources

• Bilan de 2 mois avec OpenRouter, pour 245 utilisateurs : les vrais chiffres, Nicolas Martignole · 14 September 2026 · Source documentaire
• OpenRouter, Quickstart: base URL, authentication header and endpoints · 16 September 2026 · Observation vérifiée
• OpenRouter, Provisioning API Keys: creating keys with a credit cap · 16 September 2026 · Observation vérifiée
• OpenRouter, public catalogue of models and pricing (443 models at the time of reading) · 16 September 2026 · Observation vérifiée
• Claude Max (abo prix fixe) ou API (pay-as-you-go), ce n'est pas une question de prix, Nicolas Martignole · 22 July 2026 · Source documentaire
• The State of Model Routing, NVIDIA, Cognition, OpenRouter, AI Engineer · 6 August 2026 · Source documentaire
• How We Built an AI Code Reviewer for 200 Engineers, AI Native Dev · 29 May 2026 · Source documentaire
• Anthropic's Revenue Jump, The Wealthy Bet on SpaceX · Bloomberg Tech, Bloomberg · 17 August 2026 · Source documentaire
• Chinese AI Models Gain Ground on US AI in Price and Use, Bloomberg · 22 August 2026 · Source documentaire
• Tech @ France : Comment l’État français déploie sa propre stack IA, The Product Crew · 23 April 2026 · Source documentaire
• Le vrai coût de l'IA : pourquoi le prix au token est faux, Nicolas Martignole · 11 July 2026 · Source documentaire
• Claude Code Ultimate Guide, API Gateway · 16 September 2026 · Source documentaire
• Stripe, announcement of the agreement to acquire OpenRouter · 19 August 2026 · Observation vérifiée
• OpenRouter, provider routing: zdr, data_collection, only and ignore · 17 September 2026 · Observation vérifiée
• OpenRouter, Enterprise offering: per-key caps, ZDR, regional lock-in · 17 September 2026 · Observation vérifiée
• OpenRouter, pricing: platform fees and BYOK · 17 September 2026 · Observation vérifiée
• OpenRouter, Claude Code integration guide · 17 September 2026 · Observation vérifiée
• OpenRouter and a16z, The 2025 State of AI Report · 4 December 2025 · Observation vérifiée
• Superconductor, Kimi K3 benchmark on an internal SWE-bench · 3 August 2026 · Observation vérifiée
• Inferya, cost per solved task of agent benchmarks · 21 August 2026 · Observation vérifiée
• FinOps Foundation, FinOps for AI Overview · 17 February 2026 · Observation vérifiée

“Accessed on” gives the date the source was read. A source may have been published earlier and later updated.

Update history