Users report GPT-5.5, Opus 4.7, and DeepSeek V4 limits on xhigh and 1M context

• the Opus 4.7 HN discussion says Claude Code now defaults to xhigh, while Anthropic's launch post says Opus 4.7 also ships a new tokenizer and more literal instruction following, so the model swap can change both token spend and harness behavior.
• the GPT-5.5 HN discussion caught that API access lagged the product launch, and OpenAI's launch post was later updated on April 24 to say GPT-5.5 and GPT-5.5 Pro were live in the API.
• the DeepSeek V4 discussion describes Flash as the model people could actually run, while the main HN thread on DeepSeek V4 says Pro was slower, more rate-limited, and harder to evaluate cleanly.
• DeepSeek's launch summary makes 1M tokens the headline for both V4 models, Anthropic's Opus 4.7 summary keeps pricing flat while warning about higher token counts, and OpenAI's GPT-5.5 summary pairs its own 1M context push with a staged API rollout instead of day-one full exposure.

OpenAI slipped an April 24 update into the GPT-5.5 post to confirm API access after launch-day confusion. Anthropic's own release note says older prompts can misfire because Opus 4.7 follows instructions more literally. DeepSeek's preview note gives both V4 models 1M context, but the HN thread and a Hugging Face technical writeup put the real spotlight on serving efficiency and on Flash being the practical choice.

OpenAI launched GPT-5.5 to paying ChatGPT users and Codex first, not to the API. the HN discussion called out that gap immediately, and OpenAI's launch post later added an April 24 note saying both GPT-5.5 and GPT-5.5 Pro were now available in the API.

That one-day correction mattered because the product surface and the developer surface were not the same launch. The same post that promised 1M context also said API deployments needed different safeguards, and OpenAI's system card was updated the next day with extra API-specific guardrail detail.

HN commenters also fixated on exposure paths. the main GPT-5.5 thread notes that Codex already had GPT-5.5 before the public API did, alongside reports of tighter usage limits and higher prices than earlier GPT-5.x releases.

Anthropic positioned Opus 4.7 as a straightforward GA ship: same list price as 4.6, broader availability across Claude, API, Bedrock, Vertex AI, and Foundry, plus a new xhigh effort tier in the middle of the latency curve. The buried caveat is in the same launch post: the updated tokenizer can increase token usage, and older prompts may break because 4.7 follows instructions more literally.

HN reports make that caveat less theoretical:

• the discussion summary says Claude Code now defaults to xhigh, which changes the out-of-the-box cost and latency profile.
• the main HN thread on Opus 4.7 surfaces complaints that 4.7 feels fussier on coding tasks and can refuse edits that 4.6 would make.
• the same discussion says human-readable reasoning now needs display: "summarized", another small API surface change hidden behind a version bump.
• the HN thread also quotes Anthropic's own note that prompt and harness retuning may be required.

The result is a stronger but less forgiving model. Opus 4.7 did not ship as a drop-in replacement, even though the price card barely moved.

DeepSeek shipped two very different V4 stories under one 1M-context banner. DeepSeek's preview note introduces V4-Pro at 1.6T total parameters with 49B active, and V4-Flash at 284B total with 13B active, with both exposed through new deepseek-v4-pro and deepseek-v4-flash model names. The same note also sets a retirement date, July 24, 2026, for deepseek-chat and deepseek-reasoner.

The community reaction was much less interested in the naming than in which model actually held up under load:

• the main HN thread describes Flash as the practical option so far.
• the discussion summary says Pro was heavily rate-limited and prone to timeouts.
• the same thread includes a report that Pro needed reasoning_content to avoid API errors in at least one benchmark harness.
• the discussion summary also notes that personal tests and third-party benchmarks did not clearly place V4 at the top, despite strong official numbers.

Flash getting the better early reputation is the weirdest part of the launch. Even HN commenters reported cases where Flash looked better than Pro on actual outputs, which is not how flagship-versus-cheap-tier rollouts usually read on day one.

All three launches used 1M context as a headline spec, but the rollout details point to a more practical dividing line: serving and harness constraints. the DeepSeek V4 thread is unusually explicit that Flash felt usable while Pro hit rate limits; the Opus 4.7 thread warns that xhigh defaults and tokenizer changes can raise effective cost; and the GPT-5.5 thread centers on access paths, throughput tuning, and usage caps more than on raw context length.

The vendor materials back that up with different kinds of caveats:

• Anthropic's launch post says the new tokenizer can map the same input to more tokens and that prompt behavior may change.
• OpenAI's launch post says API serving required extra safeguards, while OpenAI's system card was updated a day later for the API rollout.
• DeepSeek's preview note sells cost-effective 1M context, and Hugging Face's technical writeup argues the real novelty is the efficiency stack behind it, not the raw window size.

One concrete data point came from the GPT-5.5 thread: the HN summary quotes Codex as using traffic-shaped heuristic partitioning to lift token generation speed by more than 20%. That is a cleaner clue than any context-window slogan. In this batch of releases, the interesting story was not who printed 1M first, it was who could keep long runs affordable, available, and stable enough to matter.