OpenAI vs Anthropic: Choosing a Frontier AI Provider

A feature table comparing OpenAI and Anthropic produces a familiar shape: rows for context window, tool use, fine-tuning, image generation, a checkmark or a short phrase in each cell, and a closing “choose A if you need X, choose B if you need Y.” That framing skips the actual first question, which is not which vendor scores higher but which constraints a given organization already has, and what those constraints do to the choice before a single model benchmark enters the discussion. This page follows the wiki’s constraint-driven comparison methodology : name the constraints, separate what actually rules an option out from what is a genuine tradeoff, and compare features only for whatever survives.

This particular comparison looks different from most on this wiki, and it is worth saying so plainly rather than forcing a template. OpenAI and Anthropic are both API-first, US-headquartered frontier labs offering broadly similar products - a hosted chat completions or messages endpoint, tool use, prompt caching, a batch API. Neither requires the customer to operate physical infrastructure, so data gravity and internal capability - decisive categories in on-premise vs cloud - barely differentiate the two here: calling either API takes the same modest in-house skill, and neither vendor’s hosted service creates a training-corpus-sized dataset the customer has to move. Regulatory and jurisdictional exposure is similarly symmetric in the way that usually matters most: both companies are US entities, so the US CLOUD Act reaches both equally, and both carry the same general-purpose-AI-model provider obligations under the EU AI Act once they place a model on the EU market. An organization whose real constraint is US jurisdiction itself - the question the EU vs US AI regulation comparison is actually about - does not resolve it by picking between these two; it needs a non-US provider or a sovereign deployment, not a different frontier lab headquartered in the same country. What follows is organized around the categories that do carry real weight for this specific choice: two narrow but genuine gates, and four substantial tradeoffs.

Gates: what can rule an option out before features enter the discussion

Cloud procurement vehicle: Google Cloud is Anthropic-only; AWS and Azure now reach both vendors

This gate is narrower than it was even a few months ago, and it is worth being precise about how, because the wider claim would simply be wrong. As of September 2026, Anthropic’s Claude models are billed through the Claude API directly, through Amazon Bedrock, through Google Cloud Vertex AI, and through Claude in Microsoft Foundry on Azure - all three hyperscaler marketplaces, plus Anthropic’s own Claude Platform on AWS, a dedicated AWS Marketplace listing that bills usage in Claude Consumption Units so the whole spend shows as one line item on an AWS bill [1]. OpenAI’s models are reachable through the OpenAI API directly, through Azure OpenAI Service (its deepest, first-party home), and - since GPT-5.6 Sol, Terra, and Luna reached general availability on Amazon Bedrock on 13 July 2026 - through AWS as well [18]; GPT-6 Sol and GPT-6 Luna followed on Bedrock on 22 September 2026, the day of their release [26]. What OpenAI’s frontier models are not offered through, as of this writing, is Google Cloud: Vertex AI’s Model Garden lists Anthropic’s Claude lineup as a partner model family, but the only OpenAI entries there are the open-weight gpt-oss models, not GPT-5.6 or GPT-6 Astra [19].

That leaves one clean, narrow, genuine gate rather than the sweeping one this section used to describe: an organization whose committed-spend agreement or procurement policy is with Google Cloud specifically can procure Anthropic through Vertex AI and cannot procure OpenAI’s frontier models through Vertex AI at all, regardless of which model would otherwise be the better fit. For an AWS-committed or Azure-committed organization, that gate no longer applies the same way - both vendors are now reachable through either of those two marketplaces, which is a genuine, recent change from the “Claude means AWS, GPT means Azure” framing this page carried in earlier revisions. This is exactly the “contractual and customer-driven requirement” category the constraint-driven methodology asks to be named as what it is, not folded into a vague “compliance” row - it is just a smaller category than it used to be. See Bedrock vs Azure OpenAI for the cloud-platform mechanics in full, including the June 2026 shift that put OpenAI models on Bedrock in the first place.

Frontier cyber and bio capability: both vendors gate it behind organizational verification, symmetrically

If the specific requirement is access to the most capable available model for offensive or defensive security research - vulnerability discovery, exploit development, biosecurity work - neither vendor makes that tier available by default, and this is the one place the two companies’ September 2026 releases line up almost exactly in time and shape. Anthropic restricts Claude Mythos 5.1, released 1 September 2026, to “verified organizations” doing cyber and biological research through its Cyber Verification Program and Life Sciences Verification Program, at the same $10/$50 per MTok pricing as its general-availability sibling Claude Fable 5.1, but gated behind an approval process rather than a standard API key [2]. Two days later, OpenAI’s own GPT-6 Astra launched having crossed the “Critical” cybersecurity capability threshold in OpenAI’s Preparedness Framework - during evaluation it chained two previously unknown vulnerabilities into a working exploit without step-by-step human direction - and OpenAI is not opening that capability broadly at launch either: initial access runs through OpenAI’s Daybreak trusted-access program rather than a standard API key, with wider ChatGPT and API access following in the days after launch [3][20]. Separately from Astra’s own launch-day gating, OpenAI maintains a parallel structure for its 5.6-generation models specifically: Daybreak Blue is an alias for general-purpose models (GPT-5.6 Sol) configured with additional safeguards for defensive security work, while Daybreak Red is an alias for GPT-5.6 Cyber, a purpose-trained model restricted to authorized vulnerability research and penetration testing [21]. Neither company will simply sell an unrestricted account with this specific capability profile. For an organization that has that narrow requirement and is not already a verified counterparty at either lab, this is a real gate - but it is not a reason to prefer one vendor over the other, since both apply it, and Anthropic has one realized government-ordered shutdown to OpenAI’s zero only because Anthropic’s most capable model reached this capability threshold first, back in June - see Resilience, below. It is worth naming precisely because a comparison that only checked “does this vendor offer a cyber-capable model” (yes, both do) would miss that the actual product - unrestricted access to it - is unavailable from either.

Tradeoffs: what’s worth weighing once neither gate applies

Resilience: single-vendor dependency is not a hypothetical for either lab, and it already happened to one of them

On 12 June 2026, a US national security directive ordered Anthropic to cut off foreign access to Claude Fable 5 and Mythos 5. Because a per-user version of the order was not practical to enforce, Anthropic disabled both models for every user, worldwide, with almost no notice - Anthropic says it received the order at 5:21pm ET on a Friday [4]. It is widely described as the first time the US government restricted a specific commercial AI model, rather than the export-controlled hardware underneath it. The restriction was reversed roughly three weeks later: Anthropic announced on 30 June 2026 that it was redeploying Fable 5 from 1 July, initially capping usage at 50% of normal limits while demand caught up [5].

The lesson for this comparison is not “Anthropic is less reliable than OpenAI.” The mechanism that produced the June order - a US national security authority reaching a specific model at a specific US-headquartered lab - applies with equal legal force to OpenAI, which is also a US company; nothing in OpenAI’s structure exempts it, and this wiki is not aware of a government-ordered restriction having been issued against an OpenAI model as of this writing. But that asymmetry is a matter of timing, not of durable exposure, and it is worth being precise about why rather than leaving it as reassurance: Anthropic has one realized incident to OpenAI’s zero only because Anthropic’s most capable model crossed a national-security-relevant capability threshold first - and OpenAI’s own most capable model has, as of 1 September 2026, now crossed a comparable threshold too, by OpenAI’s own account (see Gates, above). The honest reading is closer to the opposite of a ranking: any organization whose product is wired to one vendor’s specific model, from either lab, carries this tail risk, and the risk materialized for real, for three weeks, for one of the two vendors compared on this page - and a different kind of shared-fate risk materialized for both vendors at once on the very day this page was last verified: OpenAI’s and Anthropic’s own status pages both confirmed real, multi-hour incidents that overlapped for roughly ninety minutes on the morning of 3 September 2026, OpenAI’s from 14:58 to 16:55 UTC and Anthropic’s from 13:26 to 16:23 UTC, with no shared root cause established by either company [22]. That is a different failure mode from a government order - an ordinary operational outage rather than a legal restriction - but it is the same underlying constraint category (single-vendor dependency), it hit both vendors compared on this page within the same window, and it happened on the same day this page’s other facts were verified, which makes it better evidence for the “carries this tail risk” claim above than a single one-sided anecdote would be. The concrete mitigation is identical regardless of which vendor you pick and regardless of whether the failure mode is a legal order or an ordinary outage - abstract the provider behind an interface, keep a tested fallback model, and treat “which model can I legally serve, to whom, and what happens when it’s simply down” as an architecture decision made in advance rather than an incident response made under pressure. See preparing for AI provider restrictions for the concrete pattern, the original restriction and its reversal for the full account of the June episode, and ChatGPT, Claude, and Gemini go down the same morning for the September outage.

One structural aftereffect is visible in how Anthropic ships its September 2026 lineup: Mythos 5.1 launched pre-gated to verified organizations rather than broadly available, the pattern described above under Gates - and, on the evidence of GPT-6 Astra’s own Daybreak-gated launch two days later, OpenAI’s most capable model shipped the same way. Whether either is a lasting response to a specific episode or simply how each lab would have shipped its most capability-dense model regardless is not something this page can determine.

Cost structure: comparable at the top, genuinely different at the bottom, and one promotional rate has an expiry date

Pricing below is per million tokens (MTok), current as of 25 September 2026 [1][7][23][24]. Anthropic’s per-token rate is flat across its full context window - a 900,000-token request costs the same per token as a 9,000-token one [1]. OpenAI publishes a separate, higher rate once a request crosses into “long context”: for the GPT-6 models, prompts with more than 272K input tokens are billed at 2x the input and cache rates and 1.5x the output rate for the whole request [24].

The tiers do not line up one-to-one any more: OpenAI positions GPT-6 Astra as its flagship and GPT-6 Sol as its mid tier (priced the same as GPT-5.6 Terra on input), while Anthropic’s recommended default is Opus 5.5, so the closest price comparison for Opus 5.5 is GPT-6 Sol even though the two vendors place them differently.

* There is no GPT-6 Terra; GPT-5.6 Terra remains available alongside GPT-6 Sol at the same input price. The previous-generation rows these replace - Claude Opus 5 at $5/$25 (superseded by Opus 5.5 on 22 September 2026) and GPT-5.6 Sol and Luna (superseded by GPT-6 Sol and Luna the same day) - remain available [23][24]. GPT-5.6 Sol’s $4/$20 rate is promotional pricing, published as available “at least through November 21, 2026” [7]. Anthropic went the other direction on its equivalent tier: Sonnet 5 launched at $2/$10 as an introductory rate through 31 August 2026 with a scheduled rise to $3/$15, and on 10 August 2026 Anthropic cancelled that increase and made $2/$10 the permanent standard price [8]. Neither direction should be assumed to hold for the other vendor’s tier - see Sonnet 5’s pricing becoming permanent for why an introductory rate is a marketing decision, not a cost floor, in either direction.

Three genuine, vendor-neutral differences follow from this table, not a net winner:

• At the entry tier, OpenAI is markedly cheaper. GPT-6 Luna’s $0.10/$0.50 (short context) undercuts Haiku 4.5’s $1/$5 by roughly 10x on both input and output. Against Anthropic’s default model the gap is 2x: GPT-6 Sol’s $2/$10 against Opus 5.5’s $4/$20, though OpenAI positions Sol as its mid tier rather than its flagship. For a high-volume, low-complexity workload - classification, short extraction, intent detection - that gap compounds directly into the bill.
• At long context, Anthropic’s flat pricing is structurally cheaper and more predictable. A workload that regularly pushes past OpenAI’s 272K-token threshold pays 2x the input rate and 1.5x the output rate on that vendor [24]; the same workload on Anthropic pays the single published rate regardless of context length. For agentic coding over a large repository, or RAG over long documents, this is a real architectural consideration, not a rounding difference.
• Prompt caching is structurally similar but not identical. Both vendors price a cache read at roughly 10% of the base input rate [1][7]. Anthropic’s newest models go further: cache reads on Fable 5.1 and Mythos 5.1 are priced at 2.5% of base input and on Opus 5.5 at 5%, steeper discounts than the 10% multiplier that applies to the rest of Anthropic’s lineup and to OpenAI’s cached-input pricing generally [23]. OpenAI’s GPT-6 models also bill cache writes explicitly, at 1.25x the uncached input rate [24].

Both vendors offer a Batch API at a substantial discount for asynchronous, non-time-sensitive workloads (Anthropic publishes a flat 50% off both input and output tokens across its lineup [1]); check current rates before building a cost model, and re-check the date any promotional or introductory rate expires, per the practice above.

Vendor lock-in and exit cost: neither API is portable, and the surface area differs in kind

Code written against Anthropic’s Messages API and its tool-use schema does not run against OpenAI’s Responses API (the interface GPT-6 Astra’s tool calling requires [3]) or Chat Completions without a rewrite, and the reverse is equally true - this is the same “proprietary API” lock-in mechanism covered generally in software licensing and vendor lock-in , applied to two vendors who are both, in this respect, in the same position relative to a customer. An abstraction layer or LLM gateway reduces but does not eliminate this: server-side tools (Anthropic’s web search, code execution, computer use, browser use, and memory tool; OpenAI’s web search, file search, and computer use built into the GPT-6 Astra and GPT-5.6 lines) are vendor-specific capabilities a thin routing layer does not make portable.

The two vendors differ in the shape of the lock-in they create, which is a real distinction even though neither is more locked-in than the other in the aggregate. OpenAI’s platform spans language models, embeddings (text-embedding-3-small and -large), image generation (GPT Image 2, plus GPT Image 2.5 Sunburst and Flare from 8 September 2026), speech and transcription models (including GPT-Live 1, GA 10 September 2026) [29], and until 24 September 2026 video generation (Sora-2, shut down in the API on that date with no replacement) [7][9][28] - a team that consolidates several of these onto one vendor accumulates switching cost across a wider surface than a team using only the chat endpoint. Anthropic’s product surface is narrower by design: language models plus agentic tooling, with no first-party embeddings, image, or speech models, so a team needing those must already integrate a second vendor for them regardless of which one it picks for language. Neither shape is categorically safer; one concentrates lock-in in a single, wide relationship, the other spreads it across a language-model vendor plus whatever separately handles embeddings, images, or speech.

Fine-tuning used to be a specific, real asymmetry in OpenAI’s favor, and it is now narrowing from OpenAI’s side: OpenAI is winding down its self-serve fine-tuning platform. Since 7 May 2026 organizations that had never run fine-tuning cannot create jobs, since 2 July 2026 neither can organizations without fine-tuned-model inference in the previous 60 days, and from 6 January 2027 no existing customer can create new fine-tuning jobs; existing fine-tuned models keep serving until their base model is deprecated [28]. The supported base models are also older ones (the GPT-4.1 series, and o4-mini for reinforcement fine-tuning, which shuts down 23 October 2026) [28]. Anthropic’s fine-tuning availability remains limited and enterprise-gated, largely unchanged from the position described on this page in earlier revisions, so for a new project neither vendor’s first-party API is a dependable fine-tuning path. Confirm current fine-tuning availability and terms directly with each vendor before treating this as a deciding factor - it is exactly the kind of narrow, product-surface fact that moves faster than a comparison page can track.

Trust and legal control: two different corporate structures, neither of which changes the jurisdiction question

Separate from whether either vendor is legally compellable - both are, equally, as US entities - is the distinct question of who has structural influence over a vendor’s roadmap, pricing, and model-access decisions. The two companies answer it differently.

Anthropic is a Public Benefit Corporation governed by a Board of Directors, with a Long-Term Benefit Trust (LTBT) - currently trustees Neil Buddy Shah, Richard Fontaine, and Dr. Ben Bernanke - that participates in electing that board alongside stockholders, structurally intended to weigh long-term mission considerations against short-term shareholder interest [10]. Its investor base, per its May 2026 Series H ($65 billion at a $965 billion post-money valuation), is broad: Amazon, Google, and Microsoft all participate as cloud and compute partners and investors, with Amazon investing an additional $5 billion in that round specifically, but no single entity is disclosed as a controlling shareholder [11].

OpenAI completed a different restructuring on 28 October 2025: a dual structure in which the nonprofit OpenAI Foundation holds a 26% stake in the for-profit OpenAI Group, PBC and retains the authority to remove the PBC’s board members at any time, while Microsoft holds 27% of the equity and remains OpenAI’s primary cloud and compute partner and largest single investor [12]. Both companies subsequently filed confidential IPO paperwork with the SEC in June 2026, Anthropic about a week ahead of OpenAI, both reportedly targeting valuations above $1 trillion [13].

Neither structure is a gate, and neither resolves the jurisdiction question above - an LTBT trustee and an OpenAI Foundation board seat are both, in the end, subject to the same US legal process as any other part of either company. What genuinely differs is concentration: OpenAI’s single largest external stakeholder (Microsoft, 27% equity, primary cloud partner) has a materially larger and more concentrated position than any single investor in Anthropic’s more distributed base. Whether that concentration matters to a given organization is a judgment call this page states as a fact rather than a ranking, consistent with the methodology’s instruction not to assert an unsourced verdict.

Contractual and customer-driven requirements: usually a wash, decided by your specific counterparty, not by this page

Both vendors offer enterprise plans with SSO, admin controls, and a stated default of not training on API or enterprise-plan customer data, and both maintain SOC 2 attestations and enterprise data-processing terms. As governance thresholds as you scale covers in general, a specific certification requirement is a sales-driven trigger from a specific counterparty’s procurement process, not a regulatory one - which means the deciding fact, if there is one, is what your specific customer’s specific security questionnaire names by vendor, not a general claim this page can make about either lab. Confirm current certification status, data-processing terms, and any regional data-residency options directly with each vendor’s current enterprise documentation before treating either as settled.

Model lineup, for whatever survives your constraints

Anthropic’s current lineup, per its own pricing and model documentation, current as of 25 September 2026 [1][23][25]:

Opus 5.5, released 22 September 2026 as the first model of the Claude 5.5 family, succeeded Opus 5 (the recommended default from 24 July 2026, now previous generation) and is priced below it; Anthropic has said Sonnet 5.5 and Haiku 5.5 will follow “in the coming weeks,” and neither was released as of 25 September 2026 [25]. Opus 5 had succeeded Opus 4.8 (Anthropic’s flagship from 28 May to 24 July 2026); Fable 5, Opus 5, Opus 4.8, 4.7, 4.6, and 4.5, and Sonnet 4.6 and 4.5 remain available as legacy models [23]. See Claude Opus 4.8 and Claude Sonnet 5 for the earlier releases in this sequence, and Claude Fable 5.1 reaches general availability, Mythos 5.1 stays gated for this wiki’s own coverage of the 1 September launch, including the benchmark gains Anthropic reports over Fable 5.

OpenAI’s current lineup, per its own developer documentation, current as of 25 September 2026 [7][3][9][24]:

* Promotional pricing, published as available at least through 21 November 2026 [7]. Specialized security-focused variants (GPT-5.6 Cyber, Daybreak Red/Blue) sit alongside this lineup, gated as described above.

GPT-6 Sol and GPT-6 Luna, released 22 September 2026, are reasoning models that take text and image input and are available through the Responses API, Chat Completions, and Batch; their knowledge cutoffs are 20 April and 18 May 2026 respectively [24]. They are the models that change this page’s cost comparison, since both cost half as much as their GPT-5.6 predecessors per input token and less than Anthropic’s corresponding tiers.

GPT-6 Astra began rolling out on 3 September 2026, but “GA” undersells how gated the initial release actually is: OpenAI’s own models documentation lists it as “currently rolling out to enterprise Trusted Access Program,” and OpenAI has said itself that Astra is the first OpenAI model to cross the “Critical” cybersecurity capability threshold in its Preparedness Framework, which is the reason initial access runs through the Daybreak program rather than a standard signup, with wider ChatGPT and API access following in the following days [20]. Treat its specifics as freshly published rather than battle-tested: it requires OpenAI’s Responses API for tool calling (not Chat Completions), does not support a none reasoning effort or custom temperature/top_p, and ships with a stated “misalignment monitoring” safety layer [3]. This wiki’s own independent coverage of the Critical-threshold announcement is at Astra becomes the first OpenAI model to cross the “Critical” cyber threshold ; the model-documentation facts above come directly from OpenAI’s own developer documentation and changelog. By 25 September 2026 Astra’s model page no longer carries a trusted-access rollout note, and Microsoft made Astra generally available in Microsoft Foundry on 22 September [24][27]. See OpenAI ships GPT-5.5, then GPT-5.6 for the release that preceded it.

Platform capabilities beyond language models

This is the one place a straightforward feature table earns its keep, because the asymmetry is real and stable rather than a moving target: OpenAI ships a broader platform (embeddings, image generation, speech and transcription; its video generation was shut down on 24 September 2026) alongside its language models; Anthropic does not, and instead ships agentic tooling (server-side web search, code execution, computer and browser use, a memory tool, MCP connector support, and Claude Managed Agents for stateful multi-step sessions) around its language models specifically [1][7][9].

If your requirement is confined to language and agentic tasks, this table mostly restates the lock-in-shape point made above rather than adding a new axis. If it genuinely includes embeddings, image, video, or speech, it is the one row set on this page that functions as a near-gate: Anthropic does not offer a first-party answer, so that portion of the requirement goes to OpenAI or a third vendor regardless of how the rest of this comparison resolves.

Safety posture and model philosophy

Both companies frame their alignment approach differently in public documentation. OpenAI has historically used reinforcement learning from human feedback shaped by iterative deployment and red-teaming [15]; Anthropic frames its approach around Constitutional AI, RLHF guided by an explicit written set of principles, and has generally published more externally toward safety research and model cards [16][17]. GPT-6 Astra’s stated inclusion of “misalignment monitoring” checks [3] and Anthropic’s continued system-card practice for each release (including Opus 5 [14]) suggest both labs are converging on more explicit, published safety review per release, whatever the underlying philosophical framing. Neither company’s approach is more “safe” in a way this page can rank without asserting a verdict the methodology explicitly warns against; what differs in practice is more often refusal behavior on specific edge cases than a categorical safety gap, and that is a property worth testing against your own workload rather than inferring from either company’s marketing.

What this comparison cannot resolve

• Your organization’s actual negotiated price. The rates above are public list prices; both vendors negotiate volume and enterprise terms that can shift the comparison materially, and GPT-5.6 Sol’s promotional rate specifically has a stated expiry (21 November 2026) that this page cannot predict the replacement rate for.
• Whether your specific enterprise counterparty’s security questionnaire or procurement process names one vendor by certification, region, or existing relationship. That is the counterparty’s document, not a fact this page can generalize from.
• Current fine-tuning availability and terms at Anthropic specifically, and any equivalent restrictions at OpenAI - both are the kind of product-surface fact that changes faster than a comparison page updates.
• Whether a Fable/Mythos-style government-ordered restriction could reach an OpenAI model. The legal mechanism is available against any US-headquartered lab, and OpenAI’s own Astra has now crossed a comparable capability threshold by OpenAI’s own account; whether an external order will actually be exercised against an OpenAI model, on what timeline, is not something either company’s own documentation can answer.
• Whether the 3 September 2026 OpenAI/Anthropic outage overlap shares a root cause. Neither company’s status page, nor any outlet that cross-checked them, established one; treat it as two confirmed, overlapping incidents rather than one confirmed common failure.
• How the GPT-6 and Claude 5.5 lineups settle. GPT-6 Sol and Luna and Claude Opus 5.5 were released three days before this page was last verified, and Anthropic’s Sonnet 5.5 and Haiku 5.5 have been announced but not released; treat the tier pairings above as a snapshot of a still-moving lineup, not a stable fact to build a 2027 budget on.

Further reading

• Constraint-driven comparisons : the methodology this page follows.
• Preparing for AI provider restrictions : the concrete architecture pattern for surviving a Fable/Mythos-style event, or an outage like the one on 3 September, regardless of which vendor it hits.
• Why the US restricted Anthropic’s Fable 5 and Mythos 5 and the restrictions being lifted : the full account behind the resilience section above.
• ChatGPT, Claude, and Gemini go down the same morning : the 3 September 2026 overlapping OpenAI/Anthropic outage cited in the resilience section above.
• Astra becomes the first OpenAI model to cross the “Critical” cyber threshold : the source behind GPT-6 Astra’s own Daybreak gating, cited in the gates and resilience sections above.
• Claude Fable 5.1 reaches general availability, Mythos 5.1 stays gated : this wiki’s independent coverage of the 1 September 2026 launch.
• Anthropic cancels the Sonnet 5 price rise : why an introductory rate should never be assumed permanent - or temporary - without checking.
• OpenAI ships GPT-5.5, then GPT-5.6 : the release this page’s GPT-5.6 figures build on.
• OpenAI ships GPT-6 Sol and GPT-6 Luna and Claude Opus 5.5 launches at $4/$20 : the two 22 September 2026 releases behind this page’s current pricing table.
• OpenAI’s enterprise revenue overtakes consumer : why enterprise terms now carry more commercial weight at OpenAI specifically.
• Software licensing and vendor lock-in : making an exit-cost claim concrete rather than rhetorical, for the proprietary-API side of the lock-in section above.
• The US CLOUD Act : the jurisdictional mechanism that applies symmetrically to both vendors, in full.
• EU vs US AI regulation : where jurisdiction actually is the gate, as opposed to this comparison.
• Bedrock vs Azure OpenAI : the cloud-marketplace mechanics behind the procurement-vehicle gate above, including the June 2026 shift that put OpenAI models on Bedrock.
• AI vendor selection : the general evaluation framework this page’s constraint categories fit into.
• Claude and Anthropic and OpenAI API : platform-level integration details for each.

Sources

1. Anthropic. Pricing. Model pricing, cloud platform pricing, Claude Platform on AWS pricing, and prompt-caching mechanics, fetched 3 September 2026. https://platform.claude.com/docs/en/about-claude/pricing
2. Anthropic. Models overview. Current lineup, context windows, and Mythos 5.1 limited-availability status, fetched 3 September 2026. https://platform.claude.com/docs/en/models/overview
3. OpenAI Developer Platform. API changelog, September 2026 entries (GPT-6 Astra release, Responses API requirement, “misalignment monitoring”), fetched 3 September 2026. https://developers.openai.com/api/docs/changelog
4. Multiple outlets on the 12 June 2026 restriction, collected in Why the US restricted Anthropic’s Fable 5 and Mythos 5 , including CNBC and TechCrunch, 12 June 2026.
5. Anthropic. “Redeploying Fable 5.” https://www.anthropic.com/news/redeploying-fable-5 , cited in US lifts export controls on Fable 5 and Mythos 5 .
6. Anthropic. Claude Opus 5 model page (release date, pricing, comparison table), fetched 3 September 2026. https://platform.claude.com/docs/en/models/opus-5/overview
7. OpenAI Developer Platform. Pricing and Models documentation (GPT-6 Astra, GPT-5.6 Sol/Terra/Luna, embeddings, image, video, and audio pricing), fetched 3 September 2026. https://developers.openai.com/api/docs/pricing , https://developers.openai.com/api/docs/models
8. Anthropic, Claude platform pricing documentation, note on Sonnet 5 introductory pricing becoming permanent (10 August 2026), cited in Anthropic cancels the Sonnet 5 price rise .
9. OpenAI Developer Platform. Models documentation, specialized model list (embeddings, GPT-Image-2, Sora-2, Realtime/Whisper audio family, and the Daybreak/GPT-5.6 Cyber cybersecurity category), fetched 3 September 2026. https://developers.openai.com/api/docs/models
10. Anthropic. Company page - Public Benefit Corporation status, Long-Term Benefit Trust and current trustees, board composition, fetched 3 September 2026. https://www.anthropic.com/company
11. Anthropic. “Anthropic raises Series H funding” (May 2026; investor base, Amazon’s $5B commitment, three-cloud availability). https://www.anthropic.com/news/series-h
12. Wikipedia, “OpenAI” - OpenAI Foundation / OpenAI Group PBC dual structure, equity split, and 28 October 2025 restructuring date, accessed 3 September 2026. https://en.wikipedia.org/wiki/OpenAI
13. OpenAI and Anthropic file confidential IPO paperwork : filing dates and valuation targets.
14. Anthropic. Claude Opus 5 model page - release date (24 July 2026), legacy model list, fetched 3 September 2026. https://platform.claude.com/docs/en/models/opus-5/overview
15. Ouyang, L., Wu, J., Jiang, X., et al. (2022). Training Language Models to Follow Instructions with Human Feedback. NeurIPS 2022. arXiv:2203.02155. https://arxiv.org/abs/2203.02155
16. Bai, Y., Kadavath, S., Kundu, S., et al. (2022). Constitutional AI: Harmlessness from AI Feedback. arXiv:2212.08073. https://arxiv.org/abs/2212.08073
17. Bai, Y., Jones, A., Ndousse, K., et al. (2022). Training a Helpful and Harmless Assistant with RLHF. arXiv:2204.05862. https://arxiv.org/abs/2204.05862
18. AWS. “OpenAI GPT-5.6 Sol, Terra, and Luna are now generally available on Amazon Bedrock” (13 July 2026). https://aws.amazon.com/blogs/machine-learning/openai-gpt-5-6-sol-terra-and-luna-are-now-generally-available-on-amazon-bedrock/
19. Google Cloud. Vertex AI - Use partner models (Model Garden partner-model catalog; Anthropic Claude listed, OpenAI limited to open-weight gpt-oss models), fetched 3 September 2026. https://cloud.google.com/vertex-ai/generative-ai/docs/partner-models/use-partner-models
20. This wiki, Astra becomes the first OpenAI model to cross the “Critical” cyber threshold (1-3 September 2026), sourcing OpenAI’s “Path to Astra,” CNBC, Fortune, TechCrunch, SecurityWeek, VentureBeat, NBC News, and SiliconAngle.
21. OpenAI Developer Platform. Models documentation, Cybersecurity category (Daybreak Red/Blue aliases, GPT-5.6 Cyber), fetched 3 September 2026. https://developers.openai.com/api/docs/models
22. This wiki, ChatGPT, Claude, and Gemini go down the same morning (3 September 2026), sourcing OpenAI’s and Anthropic’s own status pages directly.
23. Anthropic. Models overview - Opus 5.5 pricing, cache-read multipliers, legacy model list, fetched 25 September 2026. https://platform.claude.com/docs/en/about-claude/models/overview
24. OpenAI Developer Platform. API changelog (22 September 2026 GPT-6 Sol and Luna entry), GPT-6 Sol, GPT-6 Luna, and GPT-6 Astra model pages, and Pricing (272K long-context threshold, cache writes), fetched 25 September 2026. https://developers.openai.com/api/docs/changelog , https://developers.openai.com/api/docs/models/gpt-6-sol , https://developers.openai.com/api/docs/pricing
25. Anthropic. “Claude Opus 5.5” (22 September 2026; performance and cost claims versus Opus 5 and Fable 5.1; Sonnet 5.5 and Haiku 5.5 “in the coming weeks”). https://www.anthropic.com/claude-opus-5-5
26. AWS. “OpenAI GPT-6 Sol and GPT-6 Luna are now generally available on Amazon Bedrock” (22 September 2026). https://aws.amazon.com/about-aws/whats-new/2026/09/openai-gpt-6-sol-luna-on-amazon-bedrock/
27. Microsoft Azure Blog. “GPT-6 Astra, Sol, and Luna for production agents in Microsoft Foundry” (22 September 2026). https://azure.microsoft.com/en-us/blog/gpt-6-astra-sol-and-luna-for-production-agents-in-microsoft-foundry/
28. OpenAI Developer Platform. Deprecations - Videos API, sora-2, and sora-2-pro removed 24 September 2026 with no replacement; “Update to OpenAI’s self-serve fine-tuning” (7 May 2026, 2 July 2026, and 6 January 2027 milestones); o4-mini and ft-o4-mini shutdown 23 October 2026; fetched 25 September 2026. https://developers.openai.com/api/docs/deprecations
29. OpenAI Developer Platform. API changelog - GPT Image 2.5 Sunburst and Flare (8 September 2026), GPT-Live 1 GA and Agents API public beta (10 September 2026), and the whisper-1 / gpt-4o transcription deprecation notice (26 August 2026), fetched 25 September 2026. https://developers.openai.com/api/docs/changelog