← Blog
Field note

Claude Opus 5.5 Regression? API Changes & Migration (Nov 2026)

Thien Nguyen · Oct 6, 2026

Most Claude Opus 5.5 "regressions" are API changes, not a weaker model. Effort now defaults to medium (Opus 5 used high), thinking can't be turned off, text between tool calls arrives in thinking blocks that are empty by default, and forced tool choice returns a 400. Set effort explicitly, read blocks by type, re-run your golden set.

Updated 13 November 2026. On the OpenAI side, two "Sol" models exist and people mix them up: gpt-5.6-sol, the GPT-5.6 flagship, and gpt-6-sol, the cheaper tier below GPT-6 Astra, now followed by gpt-6.1-sol. Deadlines are forcing the move: OpenAI shuts down GPT-5.1, GPT-5.3-Codex and GPT-5.4-Nano in the API on 1 April 2027, the original GPT-5 and o3 on 11 December 2026, and Anthropic retires Claude Sonnet 4.5 on 30 November 2026. Don't just swap the model ID. Freeze a golden test set from real traffic, fix the breaking API changes, re-tune prompts and effort, shadow the new model on live traffic, then roll out behind a fallback. For most production apps that takes two to six weeks.

Model names, parameters, deadlines and prices in this playbook were re-checked on the Anthropic, OpenAI and Google model, deprecation and pricing pages on 13 November 2026. For token prices across every architecture, see our AI app development cost guide. For an eval plan that will satisfy an auditor, see LLM testing for regulated industries.

Is Claude Opus 5.5 a regression, or did the API change?

Mostly the API and its defaults changed. Anthropic lists four breaking changes and several behaviour differences from Opus 5, and each one can look like a quality drop if your code still assumes Opus 5 behaviour. Check these before you conclude the model got worse.

What users reportLikely cause on Opus 5.5Fix
"Answers are shallower than Opus 5"The default effort is medium; Opus 5 defaulted to highSet output_config.effort explicitly and run an effort sweep
"The agent goes silent between tool calls"Text between tool calls now arrives as thinking blocks, empty at the default display: "omitted"Set thinking.display to "updates" (beta) or "summarized" and render non-empty thinking blocks
"Responses are cut off" or "costs went up"Thinking is always on and counts against max_tokens; at the same effort the model also tends to think more per turn than Opus 5Raise max_tokens (start at 64k at xhigh or max), lower effort where quality holds
"My parser crashes" or "text is missing"Responses can start with thinking blocks, so content[0].text is wrongSelect content blocks by type
"It refuses things Opus 5 did"A biology classifier now runs alongside the cyber one, returning stop_reason: "refusal"Handle refusals and configure fallback (see below)
"400 errors after the switch"Disabled thinking, manual budgets, forced tool_choice, sampling parameters, prefill, old computer-use tool versions, or edited conversation historySee the parameter list in the next section

There are real capability trade-offs too, and Anthropic is open about some of them. Its launch post (22 September 2026) shows GPT-6 Astra ahead on Terminal-Bench-Science (64.6% vs 58.7%), notes that safeguard interventions lowered some benchmark scores, and says "benchmark margins have become a less reliable guide to real-world differences." That is the argument for testing on your own cases rather than trusting either a launch chart or a forum thread.

Which API parameters changed in Claude Opus 5.5?

Six request settings now return a 400 error on claude-opus-5-5: thinking disabled or with a manual budget, forced tool_choice, temperature/top_p/top_k, assistant prefill, and computer-use tool versions other than the new toolset. Effort is the only control for thinking depth. This list is from Anthropic's Opus 5.5 migration guide:

  • Thinking. Send no thinking field, or {"type": "adaptive"}, which is equivalent. {"type": "disabled"} and {"type": "enabled", "budget_tokens": N} are rejected.
  • Effort. output_config.effort accepts low, medium, high, xhigh and max. The default is medium. The effort-2025-11-24 and interleaved-thinking-2025-05-14 beta headers are no longer needed.
  • Thinking display. thinking.display defaults to "omitted". Use "summarized" for readable summaries or "updates" (beta header thinking-display-updates-2026-08-18) for progress updates only.
  • Tool choice. Only auto (default) and none. any and a named tool are rejected, including on the token-counting endpoint. Use strict: true tools or structured outputs.
  • Sampling. Omit temperature, top_p and top_k. Anthropic's deprecations page says API version 2026-10-21 removes them for every model; earlier API versions reject non-default values on Claude 4.7 and later.
  • Prefill. A final assistant turn is rejected, as on Opus 4.6 and later. Use structured outputs or system instructions.
  • Thinking blocks. Pass them back unmodified in tool loops. Blocks are tied to the model and the conversation: edit the system prompt, tools or earlier messages and the prefix check can fail, so change instructions with mid-conversation system messages instead.
  • Computer use. Declare computer_toolset_20260801; computer_20251124 works only on Amazon Bedrock.
  • Service tiers and regions. Priority Tier is not supported. US-only inference with inference_geo: "us" costs 1.1x. Fast mode is set with speed: "fast".

Prompt caching also changed in your favour: cache reads are $0.20 per million tokens, and prompts of 512 tokens or more can be cached.

Which models are being retired, and when?

These are the API deadlines most likely to hit a production app in the next six months, as listed on the vendors' deprecation pages on 13 November 2026:

Shutdown dateVendorWhat goesRecommended replacementAnnounced
30 Nov 2026 (read-only since 31 Oct)OpenAIEvals platform (dashboard and API)OpenAI recommends Promptfoo3 Jun 2026
30 Nov 2026OpenAIReusable prompts (v1/prompts) and Agent BuilderPrompts in your own code; Agents SDK or ChatGPT Workspace Agents3 Jun 2026
30 Nov 2026Anthropicclaude-sonnet-4-5-20250929claude-sonnet-5-530 Sep 2026
11 Dec 2026OpenAIOriginal gpt-5 and o3 snapshotsgpt-5.6-sol11 Jun 2026
6 Jan 2027OpenAItts-1gpt-realtime-2.1-mini1 Oct 2026
1 Feb 2027OpenAIgpt-5.5-cyber-previewThe most capable cyber model available to you5 Nov 2026
1 Apr 2027OpenAIgpt-5.1, gpt-5.3-codex, gpt-5.4-nanogpt-6-sol (Nano → gpt-6-luna)1 Oct 2026
7 May 2027 (earliest)Googlegemini-3.1-flash-litegemini-3.5-flash-liteSee Gemini deprecations page

Keep three more dates on your radar. Claude Haiku 4.5 and Claude Opus 4.5 are still active, but Anthropic only commits to keeping them "not sooner than" 15 October and 24 November 2026, so a deprecation notice could arrive at any time; Anthropic promises at least 60 days' notice before retiring a publicly released model. Gemini 3.8 Flash doesn't retire, but its price doubles on 1 January 2027, to $1.50 input and $7.50 output per million tokens. And if your regression suite still lives in OpenAI's Evals platform, export it now: it is already read-only and shuts down on 30 November.

Timeline of LLM API retirements from October 2026 to May 2027: OpenAI Evals read-only 31 Oct, Claude Sonnet 4.5 and Agent Builder 30 Nov, GPT-5 and o3 snapshots 11 Dec, TTS 6 Jan, GPT-5.1, GPT-5.3-Codex and GPT-5.4-Nano 1 Apr 2027, gemini-3.1-flash-lite 7 May 2027, with replacements
Figure 1: Timeline of LLM API retirements from October 2026 to May 2027: OpenAI Evals read-only 31 Oct, Claude Sonnet 4.5 and Agent Builder 30 Nov, GPT-5 and o3 snapshots 11 Dec, TTS 6 Jan, GPT-5.1, GPT-5.3-Codex and GPT-5.4-Nano 1 Apr 2027, gemini-3.1-flash-lite 7 May 2027, with replacements

What is Sol, the OpenAI model?

Sol is a tier name in OpenAI's model families, not a separate product. In GPT-5.6, Sol is the flagship (gpt-5.6-sol), with Terra and Luna below it. In GPT-6, the flagship is Astra, and Sol (gpt-6-sol, now gpt-6.1-sol) is the cheaper tier that OpenAI describes as near-Astra performance at lower cost.

Model IDFamily rolePrice per 1M tokens (input / output)Reasoning effort levels
gpt-6-astraGPT-6 flagship, OpenAI's recommended default$10 / $50low to max
gpt-6.1-solNear-Astra performance at lower cost$2 / $10 (cached input $0.10)low to max, no none
gpt-6-solEarlier GPT-6 Sol; replacement for GPT-5.1 and GPT-5.3-Codex$2 / $10 (cached input $0.20)none to max
gpt-6-lunaMost efficient, high-volume$0.10 / $0.50none to max
gpt-5.6-solGPT-5.6 flagship; alias gpt-5.6 routes here$4 / $20 listed (OpenAI shows promotional pricing until 30 Nov 2026)none to max
gpt-5.5Previous flagship, no shutdown date$5 / $30none to xhigh

All of these have a 1.05M-token context window (922K maximum input) and 128K maximum output on OpenAI's models pages. If you searched "Sol LLM" and meant a different open-source project with a similar name, this is not it.

How do you migrate from GPT-5.5 to GPT-5.6 Sol?

Change the model to gpt-5.6-sol, keep your GPT-5.5 reasoning effort as the baseline and test one level lower, replace prompt_cache_retention with prompt_cache_options.ttl: "30m", and re-check brevity instructions, because GPT-5.6 is more concise by default. Then run your golden set before switching traffic.

  1. Model ID. Use gpt-5.6-sol, or the gpt-5.6 alias, which routes to Sol. Pin the full ID in production so an alias change can't surprise you.
  2. Reasoning effort. GPT-5.6 adds max and supports none through max. OpenAI's GPT-5.6 Sol prompting guide says to preserve your current GPT-5.5 effort as the baseline and compare one level lower, keep high or xhigh for cases where evals show a gain, and reserve max for the hardest quality-first work.
  3. Verbosity. GPT-5.6 "tends to be more concise by default than GPT-5.5." Remove blanket "be concise" instructions that now over-shorten answers and set text.verbosity instead.
  4. Caching. Replace prompt_cache_retention with prompt_cache_options.ttl set to "30m", and review cache boundaries and cache-write billing.
  5. Tools. Expose only task-relevant tools, let the model parallelise independent reads, and test programmatic tool calling outputs separately from the final message if you use it.
  6. Cost. At listed prices, GPT-5.6 Sol is cheaper per token than GPT-5.5 ($4 / $20 vs $5 / $30). Re-check after the promotional period ends on 30 November 2026.

Before you commit, also price the jump straight to gpt-6.1-sol at $2 / $10. It skips one migration, but it requires the Responses API for tool calling and has no none effort, so it is a bigger code change. The next section covers it.

What breaks when you move to GPT-6 Sol or Claude Opus 5.5?

Both vendors made request settings that used to be harmless into hard errors. Changing the model string alone will fail in production. These are the changes their migration guides list (OpenAI's "Using GPT-6" and Anthropic's Opus 5.5 guide):

AreaGPT-6 Sol / GPT-6.1 Sol / GPT-6 LunaClaude Opus 5.5
Sampling parametersRemove temperature, top_p and top_logprobs when reasoning effort is not none (and logprobs on Chat Completions)Omit temperature, top_p and top_k; API version 2026-10-21 rejects them with a 400
Reasoning / thinkingEffort levels up to max, default medium. GPT-6.1 Sol and GPT-6 Astra have no none; use low. Replace minimal with low and compareAdaptive thinking is always on. thinking: disabled and manual budgets are rejected. Effort is the only control, and its default is medium (Opus 5 defaulted to high)
Tool callingGPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Astra need the Responses API for tools. On Chat Completions, GPT-6 Sol and Luna call functions only at effort noneForced tool choice (any or a named tool) is rejected. Use auto with strict tools or structured outputs
Response shapeReasoning settings change output length and latencyResponses can start with thinking blocks, so code that reads content[0].text breaks. Thinking blocks must be passed back unmodified in tool loops
Prefill and cachingReplace prompt_cache_retention with prompt_cache_options.ttl set to "30m"Assistant prefill is rejected; use structured outputs or system instructions
Safety and refusalsCheck refusals on your own edge casesstop_reason: "refusal" with a stop_details.category, now including "bio", with optional server-side fallback
Regional processingFast mode with EU data residency works on GPT-6.1 Sol and GPT-6 Luna, not on GPT-6 Sol or GPT-6 AstraUS-only inference costs 1.1x

Two of these cause silent regressions rather than errors. On Opus 5.5, text that the model writes between tool calls now arrives in thinking blocks, which are empty by default, so an app that streamed those progress updates goes quiet and nothing throws an exception. And on GPT-6 Sol and Opus 5.5, an unset effort parameter means medium, which can change answer depth and latency without any code change.

Monitoring dashboards for an AI application
Monitoring dashboards for an AI application

Why does Opus 5.5 refuse more, and how do you handle refusals?

Opus 5.5 runs a biology safety classifier in addition to the cybersecurity one, so it declines in more categories than Opus 5. A declined request returns HTTP 200 with stop_reason: "refusal" and a stop_details.category; handle it in code and retry on a fallback model instead of rewording the prompt until it gets through.

The categories in Anthropic's refusals and fallback guide are "cyber", "bio", "frontier_llm" (helping develop competing AI models), "reasoning_extraction" (asking the model to print its internal reasoning) and "general_harms", or null. Anthropic says benign work can trigger the cyber, bio, frontier_llm and general_harms classifiers. Anthropic does not publish an over-refusal rate comparing Opus 5.5 with Opus 5, so measure it on your own golden set.

How to handle refusals in production

  • Branch on stop_reason. A refusal can arrive before any output or mid-stream. Discard partial output in both cases.
  • Configure fallback. Set fallbacks: "default" with the server-side-fallback-2026-07-01 beta header and the API retries on the model Anthropic recommends for that category, in one call. Or name up to three fallback models, use the SDK middleware, or write your own retry. Server-side fallback is not available on the Batches API, and Haiku 5.5 has none.
  • Fix reasoning_extraction at the prompt. Prompts like "show your full chain of thought in the answer" get declined. Ask for a short explanation or the evidence instead, and set thinking.display: "summarized" if you need to see reasoning.
  • Use the verification programs for legitimate dual-use work. Security teams can apply to Anthropic's Cyber Verification Program and life-science organisations to the Life Sciences Verification Program, which Anthropic's launch post says now cover Opus 5.5.
  • Budget for billed refusals. Refusals before any output are billed in the bio, frontier_llm and reasoning_extraction categories, and every refusal counts against rate limits.
  • Track refusals as a metric. Log category, prompt type and fallback model, and add an over-refusal check to your regression suite.

Should you try to bypass Opus 5.5 refusals?

No. A classifier refusal is a policy decision, not a bug, and prompt tricks that slip past it are the wrong fix. Anthropic bills some refusals specifically "to disrupt attempts to circumvent Anthropic's safeguards at scale," a workaround that works today can break with the next classifier update, and an app built on one fails unpredictably for your users.

When a legitimate request is refused, the durable fixes are the ones above: make the system prompt state the real use case and context clearly, route the refusal to a fallback model, move dual-use work into a verification program, and keep an eval set of your legitimate edge cases so you see when a model update changes the refusal rate. If a reworded legitimate request is still declined, Anthropic's guide says to contact support with the request ID.

How do the new models compare on cost and latency?

List prices per million tokens (input / output) from the vendors' model and pricing pages on 13 November 2026. The per-1,000 columns reuse the assumptions from our cost guide: 6,000 input and 500 output tokens for a RAG answer, and 60,000 input and 3,000 output for an agent task. There is no caching, and reasoning tokens are not included.

MigrationPrice before → afterPer 1,000 RAG answersPer 1,000 agent tasksWatch out for
GPT-5.5 → GPT-5.6 Sol$5 / $30 → $4 / $20$45.00 → $34.00$390 → $300Listed price; OpenAI shows promotional pricing until 30 Nov 2026
GPT-5.5 → GPT-6.1 Sol$5 / $30 → $2 / $10$45.00 → $17.00$390 → $150Responses API required for tools; no none effort
GPT-5.1 → GPT-6 Sol$1.25 / $10 → $2 / $10$12.50 → $17.00$105 → $150Input costs 60% more. Cached input is $0.20 (Sol) or $0.10 (6.1 Sol)
GPT-5.3-Codex → GPT-6 Sol$1.75 / $14 → $2 / $10$17.50 → $17.00$147 → $150Roughly flat. Output-heavy jobs get cheaper
GPT-5.4-Nano → GPT-6 Luna$0.20 / $1.25 → $0.10 / $0.50$1.83 → $0.85$15.75 → $7.50About half the price. Check quality on your hardest cases
Claude Sonnet 4.5 → Sonnet 5.5$3 / $15 → $2 / $10$25.50 → ~$22.10$225 → ~$195The newer tokenizer uses roughly 1x to 1.35x as many tokens for the same text (we assume 1.3x), so savings are ~13%, not 33%
Claude Opus 5 → Opus 5.5$5 / $25 → $4 / $20$42.50 → $34.00$375 → $300Anthropic says Opus 5.5 costs 40% less than Opus 5 on typical workloads at default settings, partly from the lower default effort

Three cost traps sit outside the list price. First, Anthropic bills thinking tokens as output tokens, and on Opus 5.5 every request thinks, because thinking can't be turned off. Check reasoning-token usage on the OpenAI side too. A workload that previously ran without thinking can produce more output tokens even at a lower price. Second, GPT-5.5, GPT-5.6 Sol, GPT-6 Sol and GPT-6.1 Sol charge 2x input and 1.5x output for the whole request above 272K input tokens, while Claude 4.6 and later models (except Haiku 5.5) charge the standard rate across the full 1M window. Third, batch processing costs half on both vendors, so move evals and back-office jobs there.

No vendor publishes latency guarantees for these models. Anthropic says Opus 5.5 generates output more than 30% faster than Opus 5, and its fast mode (up to 2.5x output speed) costs $8 / $40. OpenAI's GPT-6.1 Sol has no none effort level, which puts a floor under its response time. Measure p50 and p95 latency on your own prompts at the effort you plan to ship. Don't trust a launch chart.

Why do model upgrades cause regressions?

Newer models are better on average and worse on some of your cases. The usual causes are:

  • Prompts tuned to the old model. Workarounds for old weaknesses, such as "always answer in JSON" or "think step by step", can confuse a model that already does those things. On Opus 5.5, asking for the reasoning in the answer can even trigger a reasoning_extraction refusal.
  • Changed defaults. Effort, verbosity and refusal behaviour shift between versions. A request that sets nothing now behaves differently.
  • Output format drift. Small changes in markdown, field order or citation style break parsers and UI.
  • Tool-use behaviour. Models differ in how readily they call tools, how many calls they make and how they handle tool errors. Agents feel this most.
  • Token counts. A new tokenizer changes context-window headroom, truncation points and cost.
  • Latency budgets. More thinking can push p95 latency past a timeout that never fired before.

Without an eval set, you find these through customer complaints. Avoiding that is the whole point of the next section.

What does an eval-driven migration plan look like?

Six steps, in this order. Two to six weeks is realistic for one production feature, depending mainly on whether you already have evals.

  1. Inventory. List every model ID, endpoint, SDK version, API version header and parameter in use. Include background jobs and scripts. Anthropic's Console can export usage broken down by API key and model, and your OpenAI usage dashboard shows which models you call.
  2. Build the golden set. Sample 200–500 real requests across your main intents, plus every past incident and known edge case. Record the current model's outputs as the baseline and label what "correct" means for each case.
  3. Fix the breaking changes. Work through the tables above. Read content blocks by type, remove sampling parameters, choose an effort level explicitly, handle refusals and move tool calls to the supported API.
  4. Tune against the set. Run the golden set on the candidate model at two or three effort levels. Remove old prompt workarounds one at a time and keep only changes that improve the score.
  5. Shadow live traffic. Send a copy of production requests to the new model without showing users the result. Compare quality scores, refusals, tool calls, cost and latency for at least a week.
  6. Canary, then cut over. Move 5%, then 25%, then 100% of traffic. Keep the old model as a fallback until its shutdown date, and keep the eval suite in CI for the next migration.
Six-step eval-driven LLM migration plan: inventory, golden set of 200 to 500 real cases, fix breaking changes, tune, shadow traffic for a week, canary 5%, 25%, 100%, plus example regression pass rules
Figure 2: Six-step eval-driven LLM migration plan: inventory, golden set of 200 to 500 real cases, fix breaking changes, tune, shadow traffic for a week, canary 5%, 25%, 100%, plus example regression pass rules

What should the regression test cover?

Use one row per risk and set the threshold before you see the results. The thresholds below are examples. Set your own from your baseline.

TestWhat it measuresExample pass rule
Task accuracyCorrect answer or action on the golden set, scored by rules, human labels or a calibrated LLM graderNo worse than baseline minus 1 point overall, and no drop on any critical intent
Format validityJSON or schema parses, required fields present100% on structured outputs
Tool-call correctnessRight tool, valid arguments, no forbidden calls, call count per taskZero forbidden calls; median call count within 20% of baseline
RefusalsOver-refusal on legitimate requests (by stop_details.category), under-refusal on red-team promptsNo new refusals on the golden set; all red-team cases blocked
GroundingAnswers supported by retrieved sources, with citations that resolveNo unsupported claims on the regulated-content subset
Latencyp50 and p95 end to end, including toolsp95 under your product timeout with margin
CostTokens and dollars per task, including reasoning tokensWithin the re-forecast budget

If your industry is regulated, keep the run results, model versions and thresholds as evidence. A model change is exactly the kind of event auditors ask about. Our regulated-industry eval plan maps these tests to HIPAA, PCI DSS and the EU AI Act, and AI agent evaluation and observability covers the tooling.

How do you roll out the new model safely in production?

Switch models through config, route by task, test the fallback path, and watch online metrics for two weeks after cutover. Five practices make that work:

  • Put the model behind a config switch, not a code change, so rollback takes seconds.
  • Route by task, not by vendor. Easy tasks can move to cheaper models such as GPT-6 Luna or Claude Haiku 5.5. Keep the expensive model for steps where your evals show it matters.
  • Configure fallback deliberately. If a refusal or outage moves a conversation to another model, test that path. Anthropic's guide warns that a fallback model will usually run without Opus 5.5's thinking blocks; the API drops blocks the target model can't read.
  • Watch online metrics for two weeks after cutover: user corrections, thumbs-down, refusals, escalations, tool errors, cost per task and p95 latency.
  • A/B test prompt changes with statistics, not intuition. A difference of two points on 50 samples is usually noise.

What is the model-migration checklist?

Fifteen checks, from inventory and golden set to refusal handling, rollback and re-forecast cost. Tick every one before you move 100% of traffic:

  • Every model ID, SDK version, API version and parameter in use is listed, including batch jobs.
  • Each model has a shutdown date and an owner in your tracker.
  • Evals are exported from any platform that is shutting down, such as OpenAI Evals before 30 November 2026.
  • A golden set of 200–500 real cases, with edge cases and past incidents, has a recorded baseline.
  • Sampling parameters, prefills, forced tool choice and disabled thinking are removed where they are rejected.
  • Response parsing reads blocks by type, and thinking blocks are passed back unmodified.
  • thinking.display is set if your UI shows progress between tool calls.
  • Effort or reasoning level is set explicitly and chosen from an eval sweep.
  • max_tokens is re-budgeted for thinking plus answer.
  • stop_reason: "refusal" is handled, with a tested fallback path.
  • Pass thresholds are written down before the candidate runs.
  • Shadow traffic has run for at least a week with quality, cost and latency compared.
  • Canary steps and a one-switch rollback are tested.
  • Token cost is re-forecast, including tokenizer changes, long-context tiers and 2027 price changes.
  • The eval suite stays in CI for the next migration.

FAQ

When does GPT-5.1 stop working in the API?

On 1 April 2027. OpenAI announced on 1 October 2026 that gpt-5.1, gpt-5.3-codex and gpt-5.4-nano will be shut down that day, with gpt-6-sol as the replacement for the first two and gpt-6-luna for Nano. Requests to a shut-down model fail.

Is GPT-5.5 being deprecated in the API?

Not as of 13 November 2026. OpenAI's deprecations page lists no shutdown date for gpt-5.5, which is priced at $5 input and $30 output per million tokens. Only gpt-5.5-cyber-preview is scheduled to shut down, on 1 February 2027. Changes to which models appear in ChatGPT are separate from API deprecations.

Is it GPT-6 Sol or GPT-5.6 Sol?

Both exist. gpt-5.6-sol is the flagship of the GPT-5.6 family and the replacement OpenAI names for the original GPT-5 and o3. gpt-6-sol is the lower-cost tier of GPT-6, below GPT-6 Astra, and the replacement for GPT-5.1 and GPT-5.3-Codex. gpt-6.1-sol is its newer version.

Should we move to GPT-6.1 Sol or Claude Opus 5.5?

Run both on your golden set and decide on the evidence. GPT-6.1 Sol costs $2 / $10 with a 1.05M context. Opus 5.5 costs $4 / $20 with a 1M context and always-on thinking. GPT-6 Astra ($10 / $50) is the comparison for the hardest work. Price per token matters less than cost per successful task, and that depends on your prompts, tools and effort level.

Can we skip evals if the new model is "better"?

No. Benchmarks average over tasks that aren't yours, and Anthropic itself notes that benchmark margins have become a less reliable guide to real-world differences. A 200-case golden set takes days to build and is the only way to know whether your app got better or worse.

How long does an LLM migration take?

Changing the model ID takes minutes. A safe migration of one production feature typically takes two to six weeks. Most of that is building the eval set, tuning prompts and effort, and running shadow traffic long enough to trust the numbers.

How can BeevR help with a model migration?

A migration is a well-bounded piece of work: inventory, evals, code changes and rollout. That makes it suitable for a fixed price. Our AI agent development work is priced by phase from $10K, with a working demo every Friday and full code ownership from day one; our pricing page explains how phases are scoped. Our open-source agent framework, Kite, ships prompt A/B testing with statistical confidence intervals on real traffic. It exists because "the new prompt feels better" is not a release criterion. If you're unsure whether your agent should be rebuilt rather than migrated, our framework guide and agent cost guide are good places to start.

BeevR is a senior, founder-led AI studio in Hanoi, Vietnam. If you have a model deadline coming up, book a 30-minute call and tell us which models you run. We'll map the breaking changes and the eval work before you commit to anything.